1 /* 2 * Copyright (c) 2001, 2012, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "code/icBuffer.hpp" 27 #include "gc_implementation/g1/bufferingOopClosure.hpp" 28 #include "gc_implementation/g1/concurrentG1Refine.hpp" 29 #include "gc_implementation/g1/concurrentG1RefineThread.hpp" 30 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp" 31 #include "gc_implementation/g1/g1AllocRegion.inline.hpp" 32 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp" 33 #include "gc_implementation/g1/g1CollectorPolicy.hpp" 34 #include "gc_implementation/g1/g1ErgoVerbose.hpp" 35 #include "gc_implementation/g1/g1EvacFailure.hpp" 36 #include "gc_implementation/g1/g1GCPhaseTimes.hpp" 37 #include "gc_implementation/g1/g1Log.hpp" 38 #include "gc_implementation/g1/g1MarkSweep.hpp" 39 #include "gc_implementation/g1/g1OopClosures.inline.hpp" 40 #include "gc_implementation/g1/g1RemSet.inline.hpp" 41 #include "gc_implementation/g1/heapRegion.inline.hpp" 42 #include "gc_implementation/g1/heapRegionRemSet.hpp" 43 #include "gc_implementation/g1/heapRegionSeq.inline.hpp" 44 #include "gc_implementation/g1/vm_operations_g1.hpp" 45 #include "gc_implementation/shared/isGCActiveMark.hpp" 46 #include "memory/gcLocker.inline.hpp" 47 #include "memory/genOopClosures.inline.hpp" 48 #include "memory/generationSpec.hpp" 49 #include "memory/referenceProcessor.hpp" 50 #include "oops/oop.inline.hpp" 51 #include "oops/oop.pcgc.inline.hpp" 52 #include "runtime/aprofiler.hpp" 53 #include "runtime/vmThread.hpp" 54 55 size_t G1CollectedHeap::_humongous_object_threshold_in_words = 0; 56 57 // turn it on so that the contents of the young list (scan-only / 58 // to-be-collected) are printed at "strategic" points before / during 59 // / after the collection --- this is useful for debugging 60 #define YOUNG_LIST_VERBOSE 0 61 // CURRENT STATUS 62 // This file is under construction. Search for "FIXME". 63 64 // INVARIANTS/NOTES 65 // 66 // All allocation activity covered by the G1CollectedHeap interface is 67 // serialized by acquiring the HeapLock. This happens in mem_allocate 68 // and allocate_new_tlab, which are the "entry" points to the 69 // allocation code from the rest of the JVM. (Note that this does not 70 // apply to TLAB allocation, which is not part of this interface: it 71 // is done by clients of this interface.) 72 73 // Notes on implementation of parallelism in different tasks. 74 // 75 // G1ParVerifyTask uses heap_region_par_iterate_chunked() for parallelism. 76 // The number of GC workers is passed to heap_region_par_iterate_chunked(). 77 // It does use run_task() which sets _n_workers in the task. 78 // G1ParTask executes g1_process_strong_roots() -> 79 // SharedHeap::process_strong_roots() which calls eventuall to 80 // CardTableModRefBS::par_non_clean_card_iterate_work() which uses 81 // SequentialSubTasksDone. SharedHeap::process_strong_roots() also 82 // directly uses SubTasksDone (_process_strong_tasks field in SharedHeap). 83 // 84 85 // Local to this file. 86 87 class RefineCardTableEntryClosure: public CardTableEntryClosure { 88 SuspendibleThreadSet* _sts; 89 G1RemSet* _g1rs; 90 ConcurrentG1Refine* _cg1r; 91 bool _concurrent; 92 public: 93 RefineCardTableEntryClosure(SuspendibleThreadSet* sts, 94 G1RemSet* g1rs, 95 ConcurrentG1Refine* cg1r) : 96 _sts(sts), _g1rs(g1rs), _cg1r(cg1r), _concurrent(true) 97 {} 98 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 99 bool oops_into_cset = _g1rs->concurrentRefineOneCard(card_ptr, worker_i, false); 100 // This path is executed by the concurrent refine or mutator threads, 101 // concurrently, and so we do not care if card_ptr contains references 102 // that point into the collection set. 103 assert(!oops_into_cset, "should be"); 104 105 if (_concurrent && _sts->should_yield()) { 106 // Caller will actually yield. 107 return false; 108 } 109 // Otherwise, we finished successfully; return true. 110 return true; 111 } 112 void set_concurrent(bool b) { _concurrent = b; } 113 }; 114 115 116 class ClearLoggedCardTableEntryClosure: public CardTableEntryClosure { 117 int _calls; 118 G1CollectedHeap* _g1h; 119 CardTableModRefBS* _ctbs; 120 int _histo[256]; 121 public: 122 ClearLoggedCardTableEntryClosure() : 123 _calls(0) 124 { 125 _g1h = G1CollectedHeap::heap(); 126 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 127 for (int i = 0; i < 256; i++) _histo[i] = 0; 128 } 129 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 130 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 131 _calls++; 132 unsigned char* ujb = (unsigned char*)card_ptr; 133 int ind = (int)(*ujb); 134 _histo[ind]++; 135 *card_ptr = -1; 136 } 137 return true; 138 } 139 int calls() { return _calls; } 140 void print_histo() { 141 gclog_or_tty->print_cr("Card table value histogram:"); 142 for (int i = 0; i < 256; i++) { 143 if (_histo[i] != 0) { 144 gclog_or_tty->print_cr(" %d: %d", i, _histo[i]); 145 } 146 } 147 } 148 }; 149 150 class RedirtyLoggedCardTableEntryClosure: public CardTableEntryClosure { 151 int _calls; 152 G1CollectedHeap* _g1h; 153 CardTableModRefBS* _ctbs; 154 public: 155 RedirtyLoggedCardTableEntryClosure() : 156 _calls(0) 157 { 158 _g1h = G1CollectedHeap::heap(); 159 _ctbs = (CardTableModRefBS*)_g1h->barrier_set(); 160 } 161 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 162 if (_g1h->is_in_reserved(_ctbs->addr_for(card_ptr))) { 163 _calls++; 164 *card_ptr = 0; 165 } 166 return true; 167 } 168 int calls() { return _calls; } 169 }; 170 171 class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure { 172 public: 173 bool do_card_ptr(jbyte* card_ptr, int worker_i) { 174 *card_ptr = CardTableModRefBS::dirty_card_val(); 175 return true; 176 } 177 }; 178 179 YoungList::YoungList(G1CollectedHeap* g1h) : 180 _g1h(g1h), _head(NULL), _length(0), _last_sampled_rs_lengths(0), 181 _survivor_head(NULL), _survivor_tail(NULL), _survivor_length(0) { 182 guarantee(check_list_empty(false), "just making sure..."); 183 } 184 185 void YoungList::push_region(HeapRegion *hr) { 186 assert(!hr->is_young(), "should not already be young"); 187 assert(hr->get_next_young_region() == NULL, "cause it should!"); 188 189 hr->set_next_young_region(_head); 190 _head = hr; 191 192 _g1h->g1_policy()->set_region_eden(hr, (int) _length); 193 ++_length; 194 } 195 196 void YoungList::add_survivor_region(HeapRegion* hr) { 197 assert(hr->is_survivor(), "should be flagged as survivor region"); 198 assert(hr->get_next_young_region() == NULL, "cause it should!"); 199 200 hr->set_next_young_region(_survivor_head); 201 if (_survivor_head == NULL) { 202 _survivor_tail = hr; 203 } 204 _survivor_head = hr; 205 ++_survivor_length; 206 } 207 208 void YoungList::empty_list(HeapRegion* list) { 209 while (list != NULL) { 210 HeapRegion* next = list->get_next_young_region(); 211 list->set_next_young_region(NULL); 212 list->uninstall_surv_rate_group(); 213 list->set_not_young(); 214 list = next; 215 } 216 } 217 218 void YoungList::empty_list() { 219 assert(check_list_well_formed(), "young list should be well formed"); 220 221 empty_list(_head); 222 _head = NULL; 223 _length = 0; 224 225 empty_list(_survivor_head); 226 _survivor_head = NULL; 227 _survivor_tail = NULL; 228 _survivor_length = 0; 229 230 _last_sampled_rs_lengths = 0; 231 232 assert(check_list_empty(false), "just making sure..."); 233 } 234 235 bool YoungList::check_list_well_formed() { 236 bool ret = true; 237 238 uint length = 0; 239 HeapRegion* curr = _head; 240 HeapRegion* last = NULL; 241 while (curr != NULL) { 242 if (!curr->is_young()) { 243 gclog_or_tty->print_cr("### YOUNG REGION "PTR_FORMAT"-"PTR_FORMAT" " 244 "incorrectly tagged (y: %d, surv: %d)", 245 curr->bottom(), curr->end(), 246 curr->is_young(), curr->is_survivor()); 247 ret = false; 248 } 249 ++length; 250 last = curr; 251 curr = curr->get_next_young_region(); 252 } 253 ret = ret && (length == _length); 254 255 if (!ret) { 256 gclog_or_tty->print_cr("### YOUNG LIST seems not well formed!"); 257 gclog_or_tty->print_cr("### list has %u entries, _length is %u", 258 length, _length); 259 } 260 261 return ret; 262 } 263 264 bool YoungList::check_list_empty(bool check_sample) { 265 bool ret = true; 266 267 if (_length != 0) { 268 gclog_or_tty->print_cr("### YOUNG LIST should have 0 length, not %u", 269 _length); 270 ret = false; 271 } 272 if (check_sample && _last_sampled_rs_lengths != 0) { 273 gclog_or_tty->print_cr("### YOUNG LIST has non-zero last sampled RS lengths"); 274 ret = false; 275 } 276 if (_head != NULL) { 277 gclog_or_tty->print_cr("### YOUNG LIST does not have a NULL head"); 278 ret = false; 279 } 280 if (!ret) { 281 gclog_or_tty->print_cr("### YOUNG LIST does not seem empty"); 282 } 283 284 return ret; 285 } 286 287 void 288 YoungList::rs_length_sampling_init() { 289 _sampled_rs_lengths = 0; 290 _curr = _head; 291 } 292 293 bool 294 YoungList::rs_length_sampling_more() { 295 return _curr != NULL; 296 } 297 298 void 299 YoungList::rs_length_sampling_next() { 300 assert( _curr != NULL, "invariant" ); 301 size_t rs_length = _curr->rem_set()->occupied(); 302 303 _sampled_rs_lengths += rs_length; 304 305 // The current region may not yet have been added to the 306 // incremental collection set (it gets added when it is 307 // retired as the current allocation region). 308 if (_curr->in_collection_set()) { 309 // Update the collection set policy information for this region 310 _g1h->g1_policy()->update_incremental_cset_info(_curr, rs_length); 311 } 312 313 _curr = _curr->get_next_young_region(); 314 if (_curr == NULL) { 315 _last_sampled_rs_lengths = _sampled_rs_lengths; 316 // gclog_or_tty->print_cr("last sampled RS lengths = %d", _last_sampled_rs_lengths); 317 } 318 } 319 320 void 321 YoungList::reset_auxilary_lists() { 322 guarantee( is_empty(), "young list should be empty" ); 323 assert(check_list_well_formed(), "young list should be well formed"); 324 325 // Add survivor regions to SurvRateGroup. 326 _g1h->g1_policy()->note_start_adding_survivor_regions(); 327 _g1h->g1_policy()->finished_recalculating_age_indexes(true /* is_survivors */); 328 329 int young_index_in_cset = 0; 330 for (HeapRegion* curr = _survivor_head; 331 curr != NULL; 332 curr = curr->get_next_young_region()) { 333 _g1h->g1_policy()->set_region_survivor(curr, young_index_in_cset); 334 335 // The region is a non-empty survivor so let's add it to 336 // the incremental collection set for the next evacuation 337 // pause. 338 _g1h->g1_policy()->add_region_to_incremental_cset_rhs(curr); 339 young_index_in_cset += 1; 340 } 341 assert((uint) young_index_in_cset == _survivor_length, "post-condition"); 342 _g1h->g1_policy()->note_stop_adding_survivor_regions(); 343 344 _head = _survivor_head; 345 _length = _survivor_length; 346 if (_survivor_head != NULL) { 347 assert(_survivor_tail != NULL, "cause it shouldn't be"); 348 assert(_survivor_length > 0, "invariant"); 349 _survivor_tail->set_next_young_region(NULL); 350 } 351 352 // Don't clear the survivor list handles until the start of 353 // the next evacuation pause - we need it in order to re-tag 354 // the survivor regions from this evacuation pause as 'young' 355 // at the start of the next. 356 357 _g1h->g1_policy()->finished_recalculating_age_indexes(false /* is_survivors */); 358 359 assert(check_list_well_formed(), "young list should be well formed"); 360 } 361 362 void YoungList::print() { 363 HeapRegion* lists[] = {_head, _survivor_head}; 364 const char* names[] = {"YOUNG", "SURVIVOR"}; 365 366 for (unsigned int list = 0; list < ARRAY_SIZE(lists); ++list) { 367 gclog_or_tty->print_cr("%s LIST CONTENTS", names[list]); 368 HeapRegion *curr = lists[list]; 369 if (curr == NULL) 370 gclog_or_tty->print_cr(" empty"); 371 while (curr != NULL) { 372 gclog_or_tty->print_cr(" "HR_FORMAT", P: "PTR_FORMAT "N: "PTR_FORMAT", age: %4d", 373 HR_FORMAT_PARAMS(curr), 374 curr->prev_top_at_mark_start(), 375 curr->next_top_at_mark_start(), 376 curr->age_in_surv_rate_group_cond()); 377 curr = curr->get_next_young_region(); 378 } 379 } 380 381 gclog_or_tty->print_cr(""); 382 } 383 384 void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr) 385 { 386 // Claim the right to put the region on the dirty cards region list 387 // by installing a self pointer. 388 HeapRegion* next = hr->get_next_dirty_cards_region(); 389 if (next == NULL) { 390 HeapRegion* res = (HeapRegion*) 391 Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(), 392 NULL); 393 if (res == NULL) { 394 HeapRegion* head; 395 do { 396 // Put the region to the dirty cards region list. 397 head = _dirty_cards_region_list; 398 next = (HeapRegion*) 399 Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head); 400 if (next == head) { 401 assert(hr->get_next_dirty_cards_region() == hr, 402 "hr->get_next_dirty_cards_region() != hr"); 403 if (next == NULL) { 404 // The last region in the list points to itself. 405 hr->set_next_dirty_cards_region(hr); 406 } else { 407 hr->set_next_dirty_cards_region(next); 408 } 409 } 410 } while (next != head); 411 } 412 } 413 } 414 415 HeapRegion* G1CollectedHeap::pop_dirty_cards_region() 416 { 417 HeapRegion* head; 418 HeapRegion* hr; 419 do { 420 head = _dirty_cards_region_list; 421 if (head == NULL) { 422 return NULL; 423 } 424 HeapRegion* new_head = head->get_next_dirty_cards_region(); 425 if (head == new_head) { 426 // The last region. 427 new_head = NULL; 428 } 429 hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list, 430 head); 431 } while (hr != head); 432 assert(hr != NULL, "invariant"); 433 hr->set_next_dirty_cards_region(NULL); 434 return hr; 435 } 436 437 void G1CollectedHeap::stop_conc_gc_threads() { 438 _cg1r->stop(); 439 _cmThread->stop(); 440 } 441 442 #ifdef ASSERT 443 // A region is added to the collection set as it is retired 444 // so an address p can point to a region which will be in the 445 // collection set but has not yet been retired. This method 446 // therefore is only accurate during a GC pause after all 447 // regions have been retired. It is used for debugging 448 // to check if an nmethod has references to objects that can 449 // be move during a partial collection. Though it can be 450 // inaccurate, it is sufficient for G1 because the conservative 451 // implementation of is_scavengable() for G1 will indicate that 452 // all nmethods must be scanned during a partial collection. 453 bool G1CollectedHeap::is_in_partial_collection(const void* p) { 454 HeapRegion* hr = heap_region_containing(p); 455 return hr != NULL && hr->in_collection_set(); 456 } 457 #endif 458 459 // Returns true if the reference points to an object that 460 // can move in an incremental collecction. 461 bool G1CollectedHeap::is_scavengable(const void* p) { 462 G1CollectedHeap* g1h = G1CollectedHeap::heap(); 463 G1CollectorPolicy* g1p = g1h->g1_policy(); 464 HeapRegion* hr = heap_region_containing(p); 465 if (hr == NULL) { 466 // perm gen (or null) 467 return false; 468 } else { 469 return !hr->isHumongous(); 470 } 471 } 472 473 void G1CollectedHeap::check_ct_logs_at_safepoint() { 474 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 475 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 476 477 // Count the dirty cards at the start. 478 CountNonCleanMemRegionClosure count1(this); 479 ct_bs->mod_card_iterate(&count1); 480 int orig_count = count1.n(); 481 482 // First clear the logged cards. 483 ClearLoggedCardTableEntryClosure clear; 484 dcqs.set_closure(&clear); 485 dcqs.apply_closure_to_all_completed_buffers(); 486 dcqs.iterate_closure_all_threads(false); 487 clear.print_histo(); 488 489 // Now ensure that there's no dirty cards. 490 CountNonCleanMemRegionClosure count2(this); 491 ct_bs->mod_card_iterate(&count2); 492 if (count2.n() != 0) { 493 gclog_or_tty->print_cr("Card table has %d entries; %d originally", 494 count2.n(), orig_count); 495 } 496 guarantee(count2.n() == 0, "Card table should be clean."); 497 498 RedirtyLoggedCardTableEntryClosure redirty; 499 JavaThread::dirty_card_queue_set().set_closure(&redirty); 500 dcqs.apply_closure_to_all_completed_buffers(); 501 dcqs.iterate_closure_all_threads(false); 502 gclog_or_tty->print_cr("Log entries = %d, dirty cards = %d.", 503 clear.calls(), orig_count); 504 guarantee(redirty.calls() == clear.calls(), 505 "Or else mechanism is broken."); 506 507 CountNonCleanMemRegionClosure count3(this); 508 ct_bs->mod_card_iterate(&count3); 509 if (count3.n() != orig_count) { 510 gclog_or_tty->print_cr("Should have restored them all: orig = %d, final = %d.", 511 orig_count, count3.n()); 512 guarantee(count3.n() >= orig_count, "Should have restored them all."); 513 } 514 515 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 516 } 517 518 // Private class members. 519 520 G1CollectedHeap* G1CollectedHeap::_g1h; 521 522 // Private methods. 523 524 HeapRegion* 525 G1CollectedHeap::new_region_try_secondary_free_list() { 526 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 527 while (!_secondary_free_list.is_empty() || free_regions_coming()) { 528 if (!_secondary_free_list.is_empty()) { 529 if (G1ConcRegionFreeingVerbose) { 530 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 531 "secondary_free_list has %u entries", 532 _secondary_free_list.length()); 533 } 534 // It looks as if there are free regions available on the 535 // secondary_free_list. Let's move them to the free_list and try 536 // again to allocate from it. 537 append_secondary_free_list(); 538 539 assert(!_free_list.is_empty(), "if the secondary_free_list was not " 540 "empty we should have moved at least one entry to the free_list"); 541 HeapRegion* res = _free_list.remove_head(); 542 if (G1ConcRegionFreeingVerbose) { 543 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 544 "allocated "HR_FORMAT" from secondary_free_list", 545 HR_FORMAT_PARAMS(res)); 546 } 547 return res; 548 } 549 550 // Wait here until we get notifed either when (a) there are no 551 // more free regions coming or (b) some regions have been moved on 552 // the secondary_free_list. 553 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 554 } 555 556 if (G1ConcRegionFreeingVerbose) { 557 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 558 "could not allocate from secondary_free_list"); 559 } 560 return NULL; 561 } 562 563 HeapRegion* G1CollectedHeap::new_region(size_t word_size, bool do_expand) { 564 assert(!isHumongous(word_size) || word_size <= HeapRegion::GrainWords, 565 "the only time we use this to allocate a humongous region is " 566 "when we are allocating a single humongous region"); 567 568 HeapRegion* res; 569 if (G1StressConcRegionFreeing) { 570 if (!_secondary_free_list.is_empty()) { 571 if (G1ConcRegionFreeingVerbose) { 572 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 573 "forced to look at the secondary_free_list"); 574 } 575 res = new_region_try_secondary_free_list(); 576 if (res != NULL) { 577 return res; 578 } 579 } 580 } 581 res = _free_list.remove_head_or_null(); 582 if (res == NULL) { 583 if (G1ConcRegionFreeingVerbose) { 584 gclog_or_tty->print_cr("G1ConcRegionFreeing [region alloc] : " 585 "res == NULL, trying the secondary_free_list"); 586 } 587 res = new_region_try_secondary_free_list(); 588 } 589 if (res == NULL && do_expand && _expand_heap_after_alloc_failure) { 590 // Currently, only attempts to allocate GC alloc regions set 591 // do_expand to true. So, we should only reach here during a 592 // safepoint. If this assumption changes we might have to 593 // reconsider the use of _expand_heap_after_alloc_failure. 594 assert(SafepointSynchronize::is_at_safepoint(), "invariant"); 595 596 ergo_verbose1(ErgoHeapSizing, 597 "attempt heap expansion", 598 ergo_format_reason("region allocation request failed") 599 ergo_format_byte("allocation request"), 600 word_size * HeapWordSize); 601 if (expand(word_size * HeapWordSize)) { 602 // Given that expand() succeeded in expanding the heap, and we 603 // always expand the heap by an amount aligned to the heap 604 // region size, the free list should in theory not be empty. So 605 // it would probably be OK to use remove_head(). But the extra 606 // check for NULL is unlikely to be a performance issue here (we 607 // just expanded the heap!) so let's just be conservative and 608 // use remove_head_or_null(). 609 res = _free_list.remove_head_or_null(); 610 } else { 611 _expand_heap_after_alloc_failure = false; 612 } 613 } 614 return res; 615 } 616 617 uint G1CollectedHeap::humongous_obj_allocate_find_first(uint num_regions, 618 size_t word_size) { 619 assert(isHumongous(word_size), "word_size should be humongous"); 620 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 621 622 uint first = G1_NULL_HRS_INDEX; 623 if (num_regions == 1) { 624 // Only one region to allocate, no need to go through the slower 625 // path. The caller will attempt the expasion if this fails, so 626 // let's not try to expand here too. 627 HeapRegion* hr = new_region(word_size, false /* do_expand */); 628 if (hr != NULL) { 629 first = hr->hrs_index(); 630 } else { 631 first = G1_NULL_HRS_INDEX; 632 } 633 } else { 634 // We can't allocate humongous regions while cleanupComplete() is 635 // running, since some of the regions we find to be empty might not 636 // yet be added to the free list and it is not straightforward to 637 // know which list they are on so that we can remove them. Note 638 // that we only need to do this if we need to allocate more than 639 // one region to satisfy the current humongous allocation 640 // request. If we are only allocating one region we use the common 641 // region allocation code (see above). 642 wait_while_free_regions_coming(); 643 append_secondary_free_list_if_not_empty_with_lock(); 644 645 if (free_regions() >= num_regions) { 646 first = _hrs.find_contiguous(num_regions); 647 if (first != G1_NULL_HRS_INDEX) { 648 for (uint i = first; i < first + num_regions; ++i) { 649 HeapRegion* hr = region_at(i); 650 assert(hr->is_empty(), "sanity"); 651 assert(is_on_master_free_list(hr), "sanity"); 652 hr->set_pending_removal(true); 653 } 654 _free_list.remove_all_pending(num_regions); 655 } 656 } 657 } 658 return first; 659 } 660 661 HeapWord* 662 G1CollectedHeap::humongous_obj_allocate_initialize_regions(uint first, 663 uint num_regions, 664 size_t word_size) { 665 assert(first != G1_NULL_HRS_INDEX, "pre-condition"); 666 assert(isHumongous(word_size), "word_size should be humongous"); 667 assert(num_regions * HeapRegion::GrainWords >= word_size, "pre-condition"); 668 669 // Index of last region in the series + 1. 670 uint last = first + num_regions; 671 672 // We need to initialize the region(s) we just discovered. This is 673 // a bit tricky given that it can happen concurrently with 674 // refinement threads refining cards on these regions and 675 // potentially wanting to refine the BOT as they are scanning 676 // those cards (this can happen shortly after a cleanup; see CR 677 // 6991377). So we have to set up the region(s) carefully and in 678 // a specific order. 679 680 // The word size sum of all the regions we will allocate. 681 size_t word_size_sum = (size_t) num_regions * HeapRegion::GrainWords; 682 assert(word_size <= word_size_sum, "sanity"); 683 684 // This will be the "starts humongous" region. 685 HeapRegion* first_hr = region_at(first); 686 // The header of the new object will be placed at the bottom of 687 // the first region. 688 HeapWord* new_obj = first_hr->bottom(); 689 // This will be the new end of the first region in the series that 690 // should also match the end of the last region in the seriers. 691 HeapWord* new_end = new_obj + word_size_sum; 692 // This will be the new top of the first region that will reflect 693 // this allocation. 694 HeapWord* new_top = new_obj + word_size; 695 696 // First, we need to zero the header of the space that we will be 697 // allocating. When we update top further down, some refinement 698 // threads might try to scan the region. By zeroing the header we 699 // ensure that any thread that will try to scan the region will 700 // come across the zero klass word and bail out. 701 // 702 // NOTE: It would not have been correct to have used 703 // CollectedHeap::fill_with_object() and make the space look like 704 // an int array. The thread that is doing the allocation will 705 // later update the object header to a potentially different array 706 // type and, for a very short period of time, the klass and length 707 // fields will be inconsistent. This could cause a refinement 708 // thread to calculate the object size incorrectly. 709 Copy::fill_to_words(new_obj, oopDesc::header_size(), 0); 710 711 // We will set up the first region as "starts humongous". This 712 // will also update the BOT covering all the regions to reflect 713 // that there is a single object that starts at the bottom of the 714 // first region. 715 first_hr->set_startsHumongous(new_top, new_end); 716 717 // Then, if there are any, we will set up the "continues 718 // humongous" regions. 719 HeapRegion* hr = NULL; 720 for (uint i = first + 1; i < last; ++i) { 721 hr = region_at(i); 722 hr->set_continuesHumongous(first_hr); 723 } 724 // If we have "continues humongous" regions (hr != NULL), then the 725 // end of the last one should match new_end. 726 assert(hr == NULL || hr->end() == new_end, "sanity"); 727 728 // Up to this point no concurrent thread would have been able to 729 // do any scanning on any region in this series. All the top 730 // fields still point to bottom, so the intersection between 731 // [bottom,top] and [card_start,card_end] will be empty. Before we 732 // update the top fields, we'll do a storestore to make sure that 733 // no thread sees the update to top before the zeroing of the 734 // object header and the BOT initialization. 735 OrderAccess::storestore(); 736 737 // Now that the BOT and the object header have been initialized, 738 // we can update top of the "starts humongous" region. 739 assert(first_hr->bottom() < new_top && new_top <= first_hr->end(), 740 "new_top should be in this region"); 741 first_hr->set_top(new_top); 742 if (_hr_printer.is_active()) { 743 HeapWord* bottom = first_hr->bottom(); 744 HeapWord* end = first_hr->orig_end(); 745 if ((first + 1) == last) { 746 // the series has a single humongous region 747 _hr_printer.alloc(G1HRPrinter::SingleHumongous, first_hr, new_top); 748 } else { 749 // the series has more than one humongous regions 750 _hr_printer.alloc(G1HRPrinter::StartsHumongous, first_hr, end); 751 } 752 } 753 754 // Now, we will update the top fields of the "continues humongous" 755 // regions. The reason we need to do this is that, otherwise, 756 // these regions would look empty and this will confuse parts of 757 // G1. For example, the code that looks for a consecutive number 758 // of empty regions will consider them empty and try to 759 // re-allocate them. We can extend is_empty() to also include 760 // !continuesHumongous(), but it is easier to just update the top 761 // fields here. The way we set top for all regions (i.e., top == 762 // end for all regions but the last one, top == new_top for the 763 // last one) is actually used when we will free up the humongous 764 // region in free_humongous_region(). 765 hr = NULL; 766 for (uint i = first + 1; i < last; ++i) { 767 hr = region_at(i); 768 if ((i + 1) == last) { 769 // last continues humongous region 770 assert(hr->bottom() < new_top && new_top <= hr->end(), 771 "new_top should fall on this region"); 772 hr->set_top(new_top); 773 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, new_top); 774 } else { 775 // not last one 776 assert(new_top > hr->end(), "new_top should be above this region"); 777 hr->set_top(hr->end()); 778 _hr_printer.alloc(G1HRPrinter::ContinuesHumongous, hr, hr->end()); 779 } 780 } 781 // If we have continues humongous regions (hr != NULL), then the 782 // end of the last one should match new_end and its top should 783 // match new_top. 784 assert(hr == NULL || 785 (hr->end() == new_end && hr->top() == new_top), "sanity"); 786 787 assert(first_hr->used() == word_size * HeapWordSize, "invariant"); 788 _summary_bytes_used += first_hr->used(); 789 _humongous_set.add(first_hr); 790 791 return new_obj; 792 } 793 794 // If could fit into free regions w/o expansion, try. 795 // Otherwise, if can expand, do so. 796 // Otherwise, if using ex regions might help, try with ex given back. 797 HeapWord* G1CollectedHeap::humongous_obj_allocate(size_t word_size) { 798 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 799 800 verify_region_sets_optional(); 801 802 size_t word_size_rounded = round_to(word_size, HeapRegion::GrainWords); 803 uint num_regions = (uint) (word_size_rounded / HeapRegion::GrainWords); 804 uint x_num = expansion_regions(); 805 uint fs = _hrs.free_suffix(); 806 uint first = humongous_obj_allocate_find_first(num_regions, word_size); 807 if (first == G1_NULL_HRS_INDEX) { 808 // The only thing we can do now is attempt expansion. 809 if (fs + x_num >= num_regions) { 810 // If the number of regions we're trying to allocate for this 811 // object is at most the number of regions in the free suffix, 812 // then the call to humongous_obj_allocate_find_first() above 813 // should have succeeded and we wouldn't be here. 814 // 815 // We should only be trying to expand when the free suffix is 816 // not sufficient for the object _and_ we have some expansion 817 // room available. 818 assert(num_regions > fs, "earlier allocation should have succeeded"); 819 820 ergo_verbose1(ErgoHeapSizing, 821 "attempt heap expansion", 822 ergo_format_reason("humongous allocation request failed") 823 ergo_format_byte("allocation request"), 824 word_size * HeapWordSize); 825 if (expand((num_regions - fs) * HeapRegion::GrainBytes)) { 826 // Even though the heap was expanded, it might not have 827 // reached the desired size. So, we cannot assume that the 828 // allocation will succeed. 829 first = humongous_obj_allocate_find_first(num_regions, word_size); 830 } 831 } 832 } 833 834 HeapWord* result = NULL; 835 if (first != G1_NULL_HRS_INDEX) { 836 result = 837 humongous_obj_allocate_initialize_regions(first, num_regions, word_size); 838 assert(result != NULL, "it should always return a valid result"); 839 840 // A successful humongous object allocation changes the used space 841 // information of the old generation so we need to recalculate the 842 // sizes and update the jstat counters here. 843 g1mm()->update_sizes(); 844 } 845 846 verify_region_sets_optional(); 847 848 return result; 849 } 850 851 HeapWord* G1CollectedHeap::allocate_new_tlab(size_t word_size) { 852 assert_heap_not_locked_and_not_at_safepoint(); 853 assert(!isHumongous(word_size), "we do not allow humongous TLABs"); 854 855 unsigned int dummy_gc_count_before; 856 return attempt_allocation(word_size, &dummy_gc_count_before); 857 } 858 859 HeapWord* 860 G1CollectedHeap::mem_allocate(size_t word_size, 861 bool* gc_overhead_limit_was_exceeded) { 862 assert_heap_not_locked_and_not_at_safepoint(); 863 864 // Loop until the allocation is satisified, or unsatisfied after GC. 865 for (int try_count = 1; /* we'll return */; try_count += 1) { 866 unsigned int gc_count_before; 867 868 HeapWord* result = NULL; 869 if (!isHumongous(word_size)) { 870 result = attempt_allocation(word_size, &gc_count_before); 871 } else { 872 result = attempt_allocation_humongous(word_size, &gc_count_before); 873 } 874 if (result != NULL) { 875 return result; 876 } 877 878 // Create the garbage collection operation... 879 VM_G1CollectForAllocation op(gc_count_before, word_size); 880 // ...and get the VM thread to execute it. 881 VMThread::execute(&op); 882 883 if (op.prologue_succeeded() && op.pause_succeeded()) { 884 // If the operation was successful we'll return the result even 885 // if it is NULL. If the allocation attempt failed immediately 886 // after a Full GC, it's unlikely we'll be able to allocate now. 887 HeapWord* result = op.result(); 888 if (result != NULL && !isHumongous(word_size)) { 889 // Allocations that take place on VM operations do not do any 890 // card dirtying and we have to do it here. We only have to do 891 // this for non-humongous allocations, though. 892 dirty_young_block(result, word_size); 893 } 894 return result; 895 } else { 896 assert(op.result() == NULL, 897 "the result should be NULL if the VM op did not succeed"); 898 } 899 900 // Give a warning if we seem to be looping forever. 901 if ((QueuedAllocationWarningCount > 0) && 902 (try_count % QueuedAllocationWarningCount == 0)) { 903 warning("G1CollectedHeap::mem_allocate retries %d times", try_count); 904 } 905 } 906 907 ShouldNotReachHere(); 908 return NULL; 909 } 910 911 HeapWord* G1CollectedHeap::attempt_allocation_slow(size_t word_size, 912 unsigned int *gc_count_before_ret) { 913 // Make sure you read the note in attempt_allocation_humongous(). 914 915 assert_heap_not_locked_and_not_at_safepoint(); 916 assert(!isHumongous(word_size), "attempt_allocation_slow() should not " 917 "be called for humongous allocation requests"); 918 919 // We should only get here after the first-level allocation attempt 920 // (attempt_allocation()) failed to allocate. 921 922 // We will loop until a) we manage to successfully perform the 923 // allocation or b) we successfully schedule a collection which 924 // fails to perform the allocation. b) is the only case when we'll 925 // return NULL. 926 HeapWord* result = NULL; 927 for (int try_count = 1; /* we'll return */; try_count += 1) { 928 bool should_try_gc; 929 unsigned int gc_count_before; 930 931 { 932 MutexLockerEx x(Heap_lock); 933 934 result = _mutator_alloc_region.attempt_allocation_locked(word_size, 935 false /* bot_updates */); 936 if (result != NULL) { 937 return result; 938 } 939 940 // If we reach here, attempt_allocation_locked() above failed to 941 // allocate a new region. So the mutator alloc region should be NULL. 942 assert(_mutator_alloc_region.get() == NULL, "only way to get here"); 943 944 if (GC_locker::is_active_and_needs_gc()) { 945 if (g1_policy()->can_expand_young_list()) { 946 // No need for an ergo verbose message here, 947 // can_expand_young_list() does this when it returns true. 948 result = _mutator_alloc_region.attempt_allocation_force(word_size, 949 false /* bot_updates */); 950 if (result != NULL) { 951 return result; 952 } 953 } 954 should_try_gc = false; 955 } else { 956 // The GCLocker may not be active but the GCLocker initiated 957 // GC may not yet have been performed (GCLocker::needs_gc() 958 // returns true). In this case we do not try this GC and 959 // wait until the GCLocker initiated GC is performed, and 960 // then retry the allocation. 961 if (GC_locker::needs_gc()) { 962 should_try_gc = false; 963 } else { 964 // Read the GC count while still holding the Heap_lock. 965 gc_count_before = total_collections(); 966 should_try_gc = true; 967 } 968 } 969 } 970 971 if (should_try_gc) { 972 bool succeeded; 973 result = do_collection_pause(word_size, gc_count_before, &succeeded); 974 if (result != NULL) { 975 assert(succeeded, "only way to get back a non-NULL result"); 976 return result; 977 } 978 979 if (succeeded) { 980 // If we get here we successfully scheduled a collection which 981 // failed to allocate. No point in trying to allocate 982 // further. We'll just return NULL. 983 MutexLockerEx x(Heap_lock); 984 *gc_count_before_ret = total_collections(); 985 return NULL; 986 } 987 } else { 988 // The GCLocker is either active or the GCLocker initiated 989 // GC has not yet been performed. Stall until it is and 990 // then retry the allocation. 991 GC_locker::stall_until_clear(); 992 } 993 994 // We can reach here if we were unsuccessul in scheduling a 995 // collection (because another thread beat us to it) or if we were 996 // stalled due to the GC locker. In either can we should retry the 997 // allocation attempt in case another thread successfully 998 // performed a collection and reclaimed enough space. We do the 999 // first attempt (without holding the Heap_lock) here and the 1000 // follow-on attempt will be at the start of the next loop 1001 // iteration (after taking the Heap_lock). 1002 result = _mutator_alloc_region.attempt_allocation(word_size, 1003 false /* bot_updates */); 1004 if (result != NULL) { 1005 return result; 1006 } 1007 1008 // Give a warning if we seem to be looping forever. 1009 if ((QueuedAllocationWarningCount > 0) && 1010 (try_count % QueuedAllocationWarningCount == 0)) { 1011 warning("G1CollectedHeap::attempt_allocation_slow() " 1012 "retries %d times", try_count); 1013 } 1014 } 1015 1016 ShouldNotReachHere(); 1017 return NULL; 1018 } 1019 1020 HeapWord* G1CollectedHeap::attempt_allocation_humongous(size_t word_size, 1021 unsigned int * gc_count_before_ret) { 1022 // The structure of this method has a lot of similarities to 1023 // attempt_allocation_slow(). The reason these two were not merged 1024 // into a single one is that such a method would require several "if 1025 // allocation is not humongous do this, otherwise do that" 1026 // conditional paths which would obscure its flow. In fact, an early 1027 // version of this code did use a unified method which was harder to 1028 // follow and, as a result, it had subtle bugs that were hard to 1029 // track down. So keeping these two methods separate allows each to 1030 // be more readable. It will be good to keep these two in sync as 1031 // much as possible. 1032 1033 assert_heap_not_locked_and_not_at_safepoint(); 1034 assert(isHumongous(word_size), "attempt_allocation_humongous() " 1035 "should only be called for humongous allocations"); 1036 1037 // Humongous objects can exhaust the heap quickly, so we should check if we 1038 // need to start a marking cycle at each humongous object allocation. We do 1039 // the check before we do the actual allocation. The reason for doing it 1040 // before the allocation is that we avoid having to keep track of the newly 1041 // allocated memory while we do a GC. 1042 if (g1_policy()->need_to_start_conc_mark("concurrent humongous allocation", 1043 word_size)) { 1044 collect(GCCause::_g1_humongous_allocation); 1045 } 1046 1047 // We will loop until a) we manage to successfully perform the 1048 // allocation or b) we successfully schedule a collection which 1049 // fails to perform the allocation. b) is the only case when we'll 1050 // return NULL. 1051 HeapWord* result = NULL; 1052 for (int try_count = 1; /* we'll return */; try_count += 1) { 1053 bool should_try_gc; 1054 unsigned int gc_count_before; 1055 1056 { 1057 MutexLockerEx x(Heap_lock); 1058 1059 // Given that humongous objects are not allocated in young 1060 // regions, we'll first try to do the allocation without doing a 1061 // collection hoping that there's enough space in the heap. 1062 result = humongous_obj_allocate(word_size); 1063 if (result != NULL) { 1064 return result; 1065 } 1066 1067 if (GC_locker::is_active_and_needs_gc()) { 1068 should_try_gc = false; 1069 } else { 1070 // The GCLocker may not be active but the GCLocker initiated 1071 // GC may not yet have been performed (GCLocker::needs_gc() 1072 // returns true). In this case we do not try this GC and 1073 // wait until the GCLocker initiated GC is performed, and 1074 // then retry the allocation. 1075 if (GC_locker::needs_gc()) { 1076 should_try_gc = false; 1077 } else { 1078 // Read the GC count while still holding the Heap_lock. 1079 gc_count_before = total_collections(); 1080 should_try_gc = true; 1081 } 1082 } 1083 } 1084 1085 if (should_try_gc) { 1086 // If we failed to allocate the humongous object, we should try to 1087 // do a collection pause (if we're allowed) in case it reclaims 1088 // enough space for the allocation to succeed after the pause. 1089 1090 bool succeeded; 1091 result = do_collection_pause(word_size, gc_count_before, &succeeded); 1092 if (result != NULL) { 1093 assert(succeeded, "only way to get back a non-NULL result"); 1094 return result; 1095 } 1096 1097 if (succeeded) { 1098 // If we get here we successfully scheduled a collection which 1099 // failed to allocate. No point in trying to allocate 1100 // further. We'll just return NULL. 1101 MutexLockerEx x(Heap_lock); 1102 *gc_count_before_ret = total_collections(); 1103 return NULL; 1104 } 1105 } else { 1106 // The GCLocker is either active or the GCLocker initiated 1107 // GC has not yet been performed. Stall until it is and 1108 // then retry the allocation. 1109 GC_locker::stall_until_clear(); 1110 } 1111 1112 // We can reach here if we were unsuccessul in scheduling a 1113 // collection (because another thread beat us to it) or if we were 1114 // stalled due to the GC locker. In either can we should retry the 1115 // allocation attempt in case another thread successfully 1116 // performed a collection and reclaimed enough space. Give a 1117 // warning if we seem to be looping forever. 1118 1119 if ((QueuedAllocationWarningCount > 0) && 1120 (try_count % QueuedAllocationWarningCount == 0)) { 1121 warning("G1CollectedHeap::attempt_allocation_humongous() " 1122 "retries %d times", try_count); 1123 } 1124 } 1125 1126 ShouldNotReachHere(); 1127 return NULL; 1128 } 1129 1130 HeapWord* G1CollectedHeap::attempt_allocation_at_safepoint(size_t word_size, 1131 bool expect_null_mutator_alloc_region) { 1132 assert_at_safepoint(true /* should_be_vm_thread */); 1133 assert(_mutator_alloc_region.get() == NULL || 1134 !expect_null_mutator_alloc_region, 1135 "the current alloc region was unexpectedly found to be non-NULL"); 1136 1137 if (!isHumongous(word_size)) { 1138 return _mutator_alloc_region.attempt_allocation_locked(word_size, 1139 false /* bot_updates */); 1140 } else { 1141 HeapWord* result = humongous_obj_allocate(word_size); 1142 if (result != NULL && g1_policy()->need_to_start_conc_mark("STW humongous allocation")) { 1143 g1_policy()->set_initiate_conc_mark_if_possible(); 1144 } 1145 return result; 1146 } 1147 1148 ShouldNotReachHere(); 1149 } 1150 1151 class PostMCRemSetClearClosure: public HeapRegionClosure { 1152 ModRefBarrierSet* _mr_bs; 1153 public: 1154 PostMCRemSetClearClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {} 1155 bool doHeapRegion(HeapRegion* r) { 1156 r->reset_gc_time_stamp(); 1157 if (r->continuesHumongous()) 1158 return false; 1159 HeapRegionRemSet* hrrs = r->rem_set(); 1160 if (hrrs != NULL) hrrs->clear(); 1161 // You might think here that we could clear just the cards 1162 // corresponding to the used region. But no: if we leave a dirty card 1163 // in a region we might allocate into, then it would prevent that card 1164 // from being enqueued, and cause it to be missed. 1165 // Re: the performance cost: we shouldn't be doing full GC anyway! 1166 _mr_bs->clear(MemRegion(r->bottom(), r->end())); 1167 return false; 1168 } 1169 }; 1170 1171 1172 class PostMCRemSetInvalidateClosure: public HeapRegionClosure { 1173 ModRefBarrierSet* _mr_bs; 1174 public: 1175 PostMCRemSetInvalidateClosure(ModRefBarrierSet* mr_bs) : _mr_bs(mr_bs) {} 1176 bool doHeapRegion(HeapRegion* r) { 1177 if (r->continuesHumongous()) return false; 1178 if (r->used_region().word_size() != 0) { 1179 _mr_bs->invalidate(r->used_region(), true /*whole heap*/); 1180 } 1181 return false; 1182 } 1183 }; 1184 1185 class RebuildRSOutOfRegionClosure: public HeapRegionClosure { 1186 G1CollectedHeap* _g1h; 1187 UpdateRSOopClosure _cl; 1188 int _worker_i; 1189 public: 1190 RebuildRSOutOfRegionClosure(G1CollectedHeap* g1, int worker_i = 0) : 1191 _cl(g1->g1_rem_set(), worker_i), 1192 _worker_i(worker_i), 1193 _g1h(g1) 1194 { } 1195 1196 bool doHeapRegion(HeapRegion* r) { 1197 if (!r->continuesHumongous()) { 1198 _cl.set_from(r); 1199 r->oop_iterate(&_cl); 1200 } 1201 return false; 1202 } 1203 }; 1204 1205 class ParRebuildRSTask: public AbstractGangTask { 1206 G1CollectedHeap* _g1; 1207 public: 1208 ParRebuildRSTask(G1CollectedHeap* g1) 1209 : AbstractGangTask("ParRebuildRSTask"), 1210 _g1(g1) 1211 { } 1212 1213 void work(uint worker_id) { 1214 RebuildRSOutOfRegionClosure rebuild_rs(_g1, worker_id); 1215 _g1->heap_region_par_iterate_chunked(&rebuild_rs, worker_id, 1216 _g1->workers()->active_workers(), 1217 HeapRegion::RebuildRSClaimValue); 1218 } 1219 }; 1220 1221 class PostCompactionPrinterClosure: public HeapRegionClosure { 1222 private: 1223 G1HRPrinter* _hr_printer; 1224 public: 1225 bool doHeapRegion(HeapRegion* hr) { 1226 assert(!hr->is_young(), "not expecting to find young regions"); 1227 // We only generate output for non-empty regions. 1228 if (!hr->is_empty()) { 1229 if (!hr->isHumongous()) { 1230 _hr_printer->post_compaction(hr, G1HRPrinter::Old); 1231 } else if (hr->startsHumongous()) { 1232 if (hr->capacity() == HeapRegion::GrainBytes) { 1233 // single humongous region 1234 _hr_printer->post_compaction(hr, G1HRPrinter::SingleHumongous); 1235 } else { 1236 _hr_printer->post_compaction(hr, G1HRPrinter::StartsHumongous); 1237 } 1238 } else { 1239 assert(hr->continuesHumongous(), "only way to get here"); 1240 _hr_printer->post_compaction(hr, G1HRPrinter::ContinuesHumongous); 1241 } 1242 } 1243 return false; 1244 } 1245 1246 PostCompactionPrinterClosure(G1HRPrinter* hr_printer) 1247 : _hr_printer(hr_printer) { } 1248 }; 1249 1250 bool G1CollectedHeap::do_collection(bool explicit_gc, 1251 bool clear_all_soft_refs, 1252 size_t word_size) { 1253 assert_at_safepoint(true /* should_be_vm_thread */); 1254 1255 if (GC_locker::check_active_before_gc()) { 1256 return false; 1257 } 1258 1259 SvcGCMarker sgcm(SvcGCMarker::FULL); 1260 ResourceMark rm; 1261 1262 print_heap_before_gc(); 1263 1264 HRSPhaseSetter x(HRSPhaseFullGC); 1265 verify_region_sets_optional(); 1266 1267 const bool do_clear_all_soft_refs = clear_all_soft_refs || 1268 collector_policy()->should_clear_all_soft_refs(); 1269 1270 ClearedAllSoftRefs casr(do_clear_all_soft_refs, collector_policy()); 1271 1272 { 1273 IsGCActiveMark x; 1274 1275 // Timing 1276 assert(gc_cause() != GCCause::_java_lang_system_gc || explicit_gc, "invariant"); 1277 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 1278 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 1279 1280 TraceTime t(GCCauseString("Full GC", gc_cause()), G1Log::fine(), true, gclog_or_tty); 1281 TraceCollectorStats tcs(g1mm()->full_collection_counters()); 1282 TraceMemoryManagerStats tms(true /* fullGC */, gc_cause()); 1283 1284 double start = os::elapsedTime(); 1285 g1_policy()->record_full_collection_start(); 1286 1287 // Note: When we have a more flexible GC logging framework that 1288 // allows us to add optional attributes to a GC log record we 1289 // could consider timing and reporting how long we wait in the 1290 // following two methods. 1291 wait_while_free_regions_coming(); 1292 // If we start the compaction before the CM threads finish 1293 // scanning the root regions we might trip them over as we'll 1294 // be moving objects / updating references. So let's wait until 1295 // they are done. By telling them to abort, they should complete 1296 // early. 1297 _cm->root_regions()->abort(); 1298 _cm->root_regions()->wait_until_scan_finished(); 1299 append_secondary_free_list_if_not_empty_with_lock(); 1300 1301 gc_prologue(true); 1302 increment_total_collections(true /* full gc */); 1303 increment_old_marking_cycles_started(); 1304 1305 size_t g1h_prev_used = used(); 1306 assert(used() == recalculate_used(), "Should be equal"); 1307 1308 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 1309 HandleMark hm; // Discard invalid handles created during verification 1310 gclog_or_tty->print(" VerifyBeforeGC:"); 1311 prepare_for_verify(); 1312 Universe::verify(/* silent */ false, 1313 /* option */ VerifyOption_G1UsePrevMarking); 1314 1315 } 1316 pre_full_gc_dump(); 1317 1318 COMPILER2_PRESENT(DerivedPointerTable::clear()); 1319 1320 // Disable discovery and empty the discovered lists 1321 // for the CM ref processor. 1322 ref_processor_cm()->disable_discovery(); 1323 ref_processor_cm()->abandon_partial_discovery(); 1324 ref_processor_cm()->verify_no_references_recorded(); 1325 1326 // Abandon current iterations of concurrent marking and concurrent 1327 // refinement, if any are in progress. We have to do this before 1328 // wait_until_scan_finished() below. 1329 concurrent_mark()->abort(); 1330 1331 // Make sure we'll choose a new allocation region afterwards. 1332 release_mutator_alloc_region(); 1333 abandon_gc_alloc_regions(); 1334 g1_rem_set()->cleanupHRRS(); 1335 1336 // We should call this after we retire any currently active alloc 1337 // regions so that all the ALLOC / RETIRE events are generated 1338 // before the start GC event. 1339 _hr_printer.start_gc(true /* full */, (size_t) total_collections()); 1340 1341 // We may have added regions to the current incremental collection 1342 // set between the last GC or pause and now. We need to clear the 1343 // incremental collection set and then start rebuilding it afresh 1344 // after this full GC. 1345 abandon_collection_set(g1_policy()->inc_cset_head()); 1346 g1_policy()->clear_incremental_cset(); 1347 g1_policy()->stop_incremental_cset_building(); 1348 1349 tear_down_region_sets(false /* free_list_only */); 1350 g1_policy()->set_gcs_are_young(true); 1351 1352 // See the comments in g1CollectedHeap.hpp and 1353 // G1CollectedHeap::ref_processing_init() about 1354 // how reference processing currently works in G1. 1355 1356 // Temporarily make discovery by the STW ref processor single threaded (non-MT). 1357 ReferenceProcessorMTDiscoveryMutator stw_rp_disc_ser(ref_processor_stw(), false); 1358 1359 // Temporarily clear the STW ref processor's _is_alive_non_header field. 1360 ReferenceProcessorIsAliveMutator stw_rp_is_alive_null(ref_processor_stw(), NULL); 1361 1362 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, true /*verify_no_refs*/); 1363 ref_processor_stw()->setup_policy(do_clear_all_soft_refs); 1364 1365 // Do collection work 1366 { 1367 HandleMark hm; // Discard invalid handles created during gc 1368 G1MarkSweep::invoke_at_safepoint(ref_processor_stw(), do_clear_all_soft_refs); 1369 } 1370 1371 assert(free_regions() == 0, "we should not have added any free regions"); 1372 rebuild_region_sets(false /* free_list_only */); 1373 1374 // Enqueue any discovered reference objects that have 1375 // not been removed from the discovered lists. 1376 ref_processor_stw()->enqueue_discovered_references(); 1377 1378 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 1379 1380 MemoryService::track_memory_usage(); 1381 1382 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 1383 HandleMark hm; // Discard invalid handles created during verification 1384 gclog_or_tty->print(" VerifyAfterGC:"); 1385 prepare_for_verify(); 1386 Universe::verify(/* silent */ false, 1387 /* option */ VerifyOption_G1UsePrevMarking); 1388 1389 } 1390 1391 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 1392 ref_processor_stw()->verify_no_references_recorded(); 1393 1394 // Note: since we've just done a full GC, concurrent 1395 // marking is no longer active. Therefore we need not 1396 // re-enable reference discovery for the CM ref processor. 1397 // That will be done at the start of the next marking cycle. 1398 assert(!ref_processor_cm()->discovery_enabled(), "Postcondition"); 1399 ref_processor_cm()->verify_no_references_recorded(); 1400 1401 reset_gc_time_stamp(); 1402 // Since everything potentially moved, we will clear all remembered 1403 // sets, and clear all cards. Later we will rebuild remebered 1404 // sets. We will also reset the GC time stamps of the regions. 1405 PostMCRemSetClearClosure rs_clear(mr_bs()); 1406 heap_region_iterate(&rs_clear); 1407 1408 // Resize the heap if necessary. 1409 resize_if_necessary_after_full_collection(explicit_gc ? 0 : word_size); 1410 1411 if (_hr_printer.is_active()) { 1412 // We should do this after we potentially resize the heap so 1413 // that all the COMMIT / UNCOMMIT events are generated before 1414 // the end GC event. 1415 1416 PostCompactionPrinterClosure cl(hr_printer()); 1417 heap_region_iterate(&cl); 1418 1419 _hr_printer.end_gc(true /* full */, (size_t) total_collections()); 1420 } 1421 1422 if (_cg1r->use_cache()) { 1423 _cg1r->clear_and_record_card_counts(); 1424 _cg1r->clear_hot_cache(); 1425 } 1426 1427 // Rebuild remembered sets of all regions. 1428 if (G1CollectedHeap::use_parallel_gc_threads()) { 1429 uint n_workers = 1430 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 1431 workers()->active_workers(), 1432 Threads::number_of_non_daemon_threads()); 1433 assert(UseDynamicNumberOfGCThreads || 1434 n_workers == workers()->total_workers(), 1435 "If not dynamic should be using all the workers"); 1436 workers()->set_active_workers(n_workers); 1437 // Set parallel threads in the heap (_n_par_threads) only 1438 // before a parallel phase and always reset it to 0 after 1439 // the phase so that the number of parallel threads does 1440 // no get carried forward to a serial phase where there 1441 // may be code that is "possibly_parallel". 1442 set_par_threads(n_workers); 1443 1444 ParRebuildRSTask rebuild_rs_task(this); 1445 assert(check_heap_region_claim_values( 1446 HeapRegion::InitialClaimValue), "sanity check"); 1447 assert(UseDynamicNumberOfGCThreads || 1448 workers()->active_workers() == workers()->total_workers(), 1449 "Unless dynamic should use total workers"); 1450 // Use the most recent number of active workers 1451 assert(workers()->active_workers() > 0, 1452 "Active workers not properly set"); 1453 set_par_threads(workers()->active_workers()); 1454 workers()->run_task(&rebuild_rs_task); 1455 set_par_threads(0); 1456 assert(check_heap_region_claim_values( 1457 HeapRegion::RebuildRSClaimValue), "sanity check"); 1458 reset_heap_region_claim_values(); 1459 } else { 1460 RebuildRSOutOfRegionClosure rebuild_rs(this); 1461 heap_region_iterate(&rebuild_rs); 1462 } 1463 1464 if (G1Log::fine()) { 1465 print_size_transition(gclog_or_tty, g1h_prev_used, used(), capacity()); 1466 } 1467 1468 if (true) { // FIXME 1469 // Ask the permanent generation to adjust size for full collections 1470 perm()->compute_new_size(); 1471 } 1472 1473 // Start a new incremental collection set for the next pause 1474 assert(g1_policy()->collection_set() == NULL, "must be"); 1475 g1_policy()->start_incremental_cset_building(); 1476 1477 // Clear the _cset_fast_test bitmap in anticipation of adding 1478 // regions to the incremental collection set for the next 1479 // evacuation pause. 1480 clear_cset_fast_test(); 1481 1482 init_mutator_alloc_region(); 1483 1484 double end = os::elapsedTime(); 1485 g1_policy()->record_full_collection_end(); 1486 1487 #ifdef TRACESPINNING 1488 ParallelTaskTerminator::print_termination_counts(); 1489 #endif 1490 1491 gc_epilogue(true); 1492 1493 // Discard all rset updates 1494 JavaThread::dirty_card_queue_set().abandon_logs(); 1495 assert(!G1DeferredRSUpdate 1496 || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any"); 1497 1498 _young_list->reset_sampled_info(); 1499 // At this point there should be no regions in the 1500 // entire heap tagged as young. 1501 assert( check_young_list_empty(true /* check_heap */), 1502 "young list should be empty at this point"); 1503 1504 // Update the number of full collections that have been completed. 1505 increment_old_marking_cycles_completed(false /* concurrent */); 1506 1507 _hrs.verify_optional(); 1508 verify_region_sets_optional(); 1509 1510 print_heap_after_gc(); 1511 1512 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 1513 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 1514 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 1515 // before any GC notifications are raised. 1516 g1mm()->update_sizes(); 1517 } 1518 1519 post_full_gc_dump(); 1520 1521 return true; 1522 } 1523 1524 void G1CollectedHeap::do_full_collection(bool clear_all_soft_refs) { 1525 // do_collection() will return whether it succeeded in performing 1526 // the GC. Currently, there is no facility on the 1527 // do_full_collection() API to notify the caller than the collection 1528 // did not succeed (e.g., because it was locked out by the GC 1529 // locker). So, right now, we'll ignore the return value. 1530 bool dummy = do_collection(true, /* explicit_gc */ 1531 clear_all_soft_refs, 1532 0 /* word_size */); 1533 } 1534 1535 // This code is mostly copied from TenuredGeneration. 1536 void 1537 G1CollectedHeap:: 1538 resize_if_necessary_after_full_collection(size_t word_size) { 1539 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, "sanity check"); 1540 1541 // Include the current allocation, if any, and bytes that will be 1542 // pre-allocated to support collections, as "used". 1543 const size_t used_after_gc = used(); 1544 const size_t capacity_after_gc = capacity(); 1545 const size_t free_after_gc = capacity_after_gc - used_after_gc; 1546 1547 // This is enforced in arguments.cpp. 1548 assert(MinHeapFreeRatio <= MaxHeapFreeRatio, 1549 "otherwise the code below doesn't make sense"); 1550 1551 // We don't have floating point command-line arguments 1552 const double minimum_free_percentage = (double) MinHeapFreeRatio / 100.0; 1553 const double maximum_used_percentage = 1.0 - minimum_free_percentage; 1554 const double maximum_free_percentage = (double) MaxHeapFreeRatio / 100.0; 1555 const double minimum_used_percentage = 1.0 - maximum_free_percentage; 1556 1557 const size_t min_heap_size = collector_policy()->min_heap_byte_size(); 1558 const size_t max_heap_size = collector_policy()->max_heap_byte_size(); 1559 1560 // We have to be careful here as these two calculations can overflow 1561 // 32-bit size_t's. 1562 double used_after_gc_d = (double) used_after_gc; 1563 double minimum_desired_capacity_d = used_after_gc_d / maximum_used_percentage; 1564 double maximum_desired_capacity_d = used_after_gc_d / minimum_used_percentage; 1565 1566 // Let's make sure that they are both under the max heap size, which 1567 // by default will make them fit into a size_t. 1568 double desired_capacity_upper_bound = (double) max_heap_size; 1569 minimum_desired_capacity_d = MIN2(minimum_desired_capacity_d, 1570 desired_capacity_upper_bound); 1571 maximum_desired_capacity_d = MIN2(maximum_desired_capacity_d, 1572 desired_capacity_upper_bound); 1573 1574 // We can now safely turn them into size_t's. 1575 size_t minimum_desired_capacity = (size_t) minimum_desired_capacity_d; 1576 size_t maximum_desired_capacity = (size_t) maximum_desired_capacity_d; 1577 1578 // This assert only makes sense here, before we adjust them 1579 // with respect to the min and max heap size. 1580 assert(minimum_desired_capacity <= maximum_desired_capacity, 1581 err_msg("minimum_desired_capacity = "SIZE_FORMAT", " 1582 "maximum_desired_capacity = "SIZE_FORMAT, 1583 minimum_desired_capacity, maximum_desired_capacity)); 1584 1585 // Should not be greater than the heap max size. No need to adjust 1586 // it with respect to the heap min size as it's a lower bound (i.e., 1587 // we'll try to make the capacity larger than it, not smaller). 1588 minimum_desired_capacity = MIN2(minimum_desired_capacity, max_heap_size); 1589 // Should not be less than the heap min size. No need to adjust it 1590 // with respect to the heap max size as it's an upper bound (i.e., 1591 // we'll try to make the capacity smaller than it, not greater). 1592 maximum_desired_capacity = MAX2(maximum_desired_capacity, min_heap_size); 1593 1594 if (capacity_after_gc < minimum_desired_capacity) { 1595 // Don't expand unless it's significant 1596 size_t expand_bytes = minimum_desired_capacity - capacity_after_gc; 1597 ergo_verbose4(ErgoHeapSizing, 1598 "attempt heap expansion", 1599 ergo_format_reason("capacity lower than " 1600 "min desired capacity after Full GC") 1601 ergo_format_byte("capacity") 1602 ergo_format_byte("occupancy") 1603 ergo_format_byte_perc("min desired capacity"), 1604 capacity_after_gc, used_after_gc, 1605 minimum_desired_capacity, (double) MinHeapFreeRatio); 1606 expand(expand_bytes); 1607 1608 // No expansion, now see if we want to shrink 1609 } else if (capacity_after_gc > maximum_desired_capacity) { 1610 // Capacity too large, compute shrinking size 1611 size_t shrink_bytes = capacity_after_gc - maximum_desired_capacity; 1612 ergo_verbose4(ErgoHeapSizing, 1613 "attempt heap shrinking", 1614 ergo_format_reason("capacity higher than " 1615 "max desired capacity after Full GC") 1616 ergo_format_byte("capacity") 1617 ergo_format_byte("occupancy") 1618 ergo_format_byte_perc("max desired capacity"), 1619 capacity_after_gc, used_after_gc, 1620 maximum_desired_capacity, (double) MaxHeapFreeRatio); 1621 shrink(shrink_bytes); 1622 } 1623 } 1624 1625 1626 HeapWord* 1627 G1CollectedHeap::satisfy_failed_allocation(size_t word_size, 1628 bool* succeeded) { 1629 assert_at_safepoint(true /* should_be_vm_thread */); 1630 1631 *succeeded = true; 1632 // Let's attempt the allocation first. 1633 HeapWord* result = 1634 attempt_allocation_at_safepoint(word_size, 1635 false /* expect_null_mutator_alloc_region */); 1636 if (result != NULL) { 1637 assert(*succeeded, "sanity"); 1638 return result; 1639 } 1640 1641 // In a G1 heap, we're supposed to keep allocation from failing by 1642 // incremental pauses. Therefore, at least for now, we'll favor 1643 // expansion over collection. (This might change in the future if we can 1644 // do something smarter than full collection to satisfy a failed alloc.) 1645 result = expand_and_allocate(word_size); 1646 if (result != NULL) { 1647 assert(*succeeded, "sanity"); 1648 return result; 1649 } 1650 1651 // Expansion didn't work, we'll try to do a Full GC. 1652 bool gc_succeeded = do_collection(false, /* explicit_gc */ 1653 false, /* clear_all_soft_refs */ 1654 word_size); 1655 if (!gc_succeeded) { 1656 *succeeded = false; 1657 return NULL; 1658 } 1659 1660 // Retry the allocation 1661 result = attempt_allocation_at_safepoint(word_size, 1662 true /* expect_null_mutator_alloc_region */); 1663 if (result != NULL) { 1664 assert(*succeeded, "sanity"); 1665 return result; 1666 } 1667 1668 // Then, try a Full GC that will collect all soft references. 1669 gc_succeeded = do_collection(false, /* explicit_gc */ 1670 true, /* clear_all_soft_refs */ 1671 word_size); 1672 if (!gc_succeeded) { 1673 *succeeded = false; 1674 return NULL; 1675 } 1676 1677 // Retry the allocation once more 1678 result = attempt_allocation_at_safepoint(word_size, 1679 true /* expect_null_mutator_alloc_region */); 1680 if (result != NULL) { 1681 assert(*succeeded, "sanity"); 1682 return result; 1683 } 1684 1685 assert(!collector_policy()->should_clear_all_soft_refs(), 1686 "Flag should have been handled and cleared prior to this point"); 1687 1688 // What else? We might try synchronous finalization later. If the total 1689 // space available is large enough for the allocation, then a more 1690 // complete compaction phase than we've tried so far might be 1691 // appropriate. 1692 assert(*succeeded, "sanity"); 1693 return NULL; 1694 } 1695 1696 // Attempting to expand the heap sufficiently 1697 // to support an allocation of the given "word_size". If 1698 // successful, perform the allocation and return the address of the 1699 // allocated block, or else "NULL". 1700 1701 HeapWord* G1CollectedHeap::expand_and_allocate(size_t word_size) { 1702 assert_at_safepoint(true /* should_be_vm_thread */); 1703 1704 verify_region_sets_optional(); 1705 1706 size_t expand_bytes = MAX2(word_size * HeapWordSize, MinHeapDeltaBytes); 1707 ergo_verbose1(ErgoHeapSizing, 1708 "attempt heap expansion", 1709 ergo_format_reason("allocation request failed") 1710 ergo_format_byte("allocation request"), 1711 word_size * HeapWordSize); 1712 if (expand(expand_bytes)) { 1713 _hrs.verify_optional(); 1714 verify_region_sets_optional(); 1715 return attempt_allocation_at_safepoint(word_size, 1716 false /* expect_null_mutator_alloc_region */); 1717 } 1718 return NULL; 1719 } 1720 1721 void G1CollectedHeap::update_committed_space(HeapWord* old_end, 1722 HeapWord* new_end) { 1723 assert(old_end != new_end, "don't call this otherwise"); 1724 assert((HeapWord*) _g1_storage.high() == new_end, "invariant"); 1725 1726 // Update the committed mem region. 1727 _g1_committed.set_end(new_end); 1728 // Tell the card table about the update. 1729 Universe::heap()->barrier_set()->resize_covered_region(_g1_committed); 1730 // Tell the BOT about the update. 1731 _bot_shared->resize(_g1_committed.word_size()); 1732 } 1733 1734 bool G1CollectedHeap::expand(size_t expand_bytes) { 1735 size_t old_mem_size = _g1_storage.committed_size(); 1736 size_t aligned_expand_bytes = ReservedSpace::page_align_size_up(expand_bytes); 1737 aligned_expand_bytes = align_size_up(aligned_expand_bytes, 1738 HeapRegion::GrainBytes); 1739 ergo_verbose2(ErgoHeapSizing, 1740 "expand the heap", 1741 ergo_format_byte("requested expansion amount") 1742 ergo_format_byte("attempted expansion amount"), 1743 expand_bytes, aligned_expand_bytes); 1744 1745 // First commit the memory. 1746 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1747 bool successful = _g1_storage.expand_by(aligned_expand_bytes); 1748 if (successful) { 1749 // Then propagate this update to the necessary data structures. 1750 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1751 update_committed_space(old_end, new_end); 1752 1753 FreeRegionList expansion_list("Local Expansion List"); 1754 MemRegion mr = _hrs.expand_by(old_end, new_end, &expansion_list); 1755 assert(mr.start() == old_end, "post-condition"); 1756 // mr might be a smaller region than what was requested if 1757 // expand_by() was unable to allocate the HeapRegion instances 1758 assert(mr.end() <= new_end, "post-condition"); 1759 1760 size_t actual_expand_bytes = mr.byte_size(); 1761 assert(actual_expand_bytes <= aligned_expand_bytes, "post-condition"); 1762 assert(actual_expand_bytes == expansion_list.total_capacity_bytes(), 1763 "post-condition"); 1764 if (actual_expand_bytes < aligned_expand_bytes) { 1765 // We could not expand _hrs to the desired size. In this case we 1766 // need to shrink the committed space accordingly. 1767 assert(mr.end() < new_end, "invariant"); 1768 1769 size_t diff_bytes = aligned_expand_bytes - actual_expand_bytes; 1770 // First uncommit the memory. 1771 _g1_storage.shrink_by(diff_bytes); 1772 // Then propagate this update to the necessary data structures. 1773 update_committed_space(new_end, mr.end()); 1774 } 1775 _free_list.add_as_tail(&expansion_list); 1776 1777 if (_hr_printer.is_active()) { 1778 HeapWord* curr = mr.start(); 1779 while (curr < mr.end()) { 1780 HeapWord* curr_end = curr + HeapRegion::GrainWords; 1781 _hr_printer.commit(curr, curr_end); 1782 curr = curr_end; 1783 } 1784 assert(curr == mr.end(), "post-condition"); 1785 } 1786 g1_policy()->record_new_heap_size(n_regions()); 1787 } else { 1788 ergo_verbose0(ErgoHeapSizing, 1789 "did not expand the heap", 1790 ergo_format_reason("heap expansion operation failed")); 1791 // The expansion of the virtual storage space was unsuccessful. 1792 // Let's see if it was because we ran out of swap. 1793 if (G1ExitOnExpansionFailure && 1794 _g1_storage.uncommitted_size() >= aligned_expand_bytes) { 1795 // We had head room... 1796 vm_exit_out_of_memory(aligned_expand_bytes, "G1 heap expansion"); 1797 } 1798 } 1799 return successful; 1800 } 1801 1802 void G1CollectedHeap::shrink_helper(size_t shrink_bytes) { 1803 size_t old_mem_size = _g1_storage.committed_size(); 1804 size_t aligned_shrink_bytes = 1805 ReservedSpace::page_align_size_down(shrink_bytes); 1806 aligned_shrink_bytes = align_size_down(aligned_shrink_bytes, 1807 HeapRegion::GrainBytes); 1808 uint num_regions_deleted = 0; 1809 MemRegion mr = _hrs.shrink_by(aligned_shrink_bytes, &num_regions_deleted); 1810 HeapWord* old_end = (HeapWord*) _g1_storage.high(); 1811 assert(mr.end() == old_end, "post-condition"); 1812 1813 ergo_verbose3(ErgoHeapSizing, 1814 "shrink the heap", 1815 ergo_format_byte("requested shrinking amount") 1816 ergo_format_byte("aligned shrinking amount") 1817 ergo_format_byte("attempted shrinking amount"), 1818 shrink_bytes, aligned_shrink_bytes, mr.byte_size()); 1819 if (mr.byte_size() > 0) { 1820 if (_hr_printer.is_active()) { 1821 HeapWord* curr = mr.end(); 1822 while (curr > mr.start()) { 1823 HeapWord* curr_end = curr; 1824 curr -= HeapRegion::GrainWords; 1825 _hr_printer.uncommit(curr, curr_end); 1826 } 1827 assert(curr == mr.start(), "post-condition"); 1828 } 1829 1830 _g1_storage.shrink_by(mr.byte_size()); 1831 HeapWord* new_end = (HeapWord*) _g1_storage.high(); 1832 assert(mr.start() == new_end, "post-condition"); 1833 1834 _expansion_regions += num_regions_deleted; 1835 update_committed_space(old_end, new_end); 1836 HeapRegionRemSet::shrink_heap(n_regions()); 1837 g1_policy()->record_new_heap_size(n_regions()); 1838 } else { 1839 ergo_verbose0(ErgoHeapSizing, 1840 "did not shrink the heap", 1841 ergo_format_reason("heap shrinking operation failed")); 1842 } 1843 } 1844 1845 void G1CollectedHeap::shrink(size_t shrink_bytes) { 1846 verify_region_sets_optional(); 1847 1848 // We should only reach here at the end of a Full GC which means we 1849 // should not not be holding to any GC alloc regions. The method 1850 // below will make sure of that and do any remaining clean up. 1851 abandon_gc_alloc_regions(); 1852 1853 // Instead of tearing down / rebuilding the free lists here, we 1854 // could instead use the remove_all_pending() method on free_list to 1855 // remove only the ones that we need to remove. 1856 tear_down_region_sets(true /* free_list_only */); 1857 shrink_helper(shrink_bytes); 1858 rebuild_region_sets(true /* free_list_only */); 1859 1860 _hrs.verify_optional(); 1861 verify_region_sets_optional(); 1862 } 1863 1864 // Public methods. 1865 1866 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away 1867 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list 1868 #endif // _MSC_VER 1869 1870 1871 G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) : 1872 SharedHeap(policy_), 1873 _g1_policy(policy_), 1874 _dirty_card_queue_set(false), 1875 _into_cset_dirty_card_queue_set(false), 1876 _is_alive_closure_cm(this), 1877 _is_alive_closure_stw(this), 1878 _ref_processor_cm(NULL), 1879 _ref_processor_stw(NULL), 1880 _process_strong_tasks(new SubTasksDone(G1H_PS_NumElements)), 1881 _bot_shared(NULL), 1882 _objs_with_preserved_marks(NULL), _preserved_marks_of_objs(NULL), 1883 _evac_failure_scan_stack(NULL) , 1884 _mark_in_progress(false), 1885 _cg1r(NULL), _summary_bytes_used(0), 1886 _g1mm(NULL), 1887 _refine_cte_cl(NULL), 1888 _full_collection(false), 1889 _free_list("Master Free List"), 1890 _secondary_free_list("Secondary Free List"), 1891 _old_set("Old Set"), 1892 _humongous_set("Master Humongous Set"), 1893 _free_regions_coming(false), 1894 _young_list(new YoungList(this)), 1895 _gc_time_stamp(0), 1896 _retained_old_gc_alloc_region(NULL), 1897 _expand_heap_after_alloc_failure(true), 1898 _surviving_young_words(NULL), 1899 _old_marking_cycles_started(0), 1900 _old_marking_cycles_completed(0), 1901 _in_cset_fast_test(NULL), 1902 _in_cset_fast_test_base(NULL), 1903 _dirty_cards_region_list(NULL), 1904 _worker_cset_start_region(NULL), 1905 _worker_cset_start_region_time_stamp(NULL) { 1906 _g1h = this; // To catch bugs. 1907 if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) { 1908 vm_exit_during_initialization("Failed necessary allocation."); 1909 } 1910 1911 _humongous_object_threshold_in_words = HeapRegion::GrainWords / 2; 1912 1913 int n_queues = MAX2((int)ParallelGCThreads, 1); 1914 _task_queues = new RefToScanQueueSet(n_queues); 1915 1916 int n_rem_sets = HeapRegionRemSet::num_par_rem_sets(); 1917 assert(n_rem_sets > 0, "Invariant."); 1918 1919 HeapRegionRemSetIterator** iter_arr = 1920 NEW_C_HEAP_ARRAY(HeapRegionRemSetIterator*, n_queues, mtGC); 1921 for (int i = 0; i < n_queues; i++) { 1922 iter_arr[i] = new HeapRegionRemSetIterator(); 1923 } 1924 _rem_set_iterator = iter_arr; 1925 1926 _worker_cset_start_region = NEW_C_HEAP_ARRAY(HeapRegion*, n_queues, mtGC); 1927 _worker_cset_start_region_time_stamp = NEW_C_HEAP_ARRAY(unsigned int, n_queues, mtGC); 1928 1929 for (int i = 0; i < n_queues; i++) { 1930 RefToScanQueue* q = new RefToScanQueue(); 1931 q->initialize(); 1932 _task_queues->register_queue(i, q); 1933 } 1934 1935 clear_cset_start_regions(); 1936 1937 guarantee(_task_queues != NULL, "task_queues allocation failure."); 1938 } 1939 1940 jint G1CollectedHeap::initialize() { 1941 CollectedHeap::pre_initialize(); 1942 os::enable_vtime(); 1943 1944 G1Log::init(); 1945 1946 // Necessary to satisfy locking discipline assertions. 1947 1948 MutexLocker x(Heap_lock); 1949 1950 // We have to initialize the printer before committing the heap, as 1951 // it will be used then. 1952 _hr_printer.set_active(G1PrintHeapRegions); 1953 1954 // While there are no constraints in the GC code that HeapWordSize 1955 // be any particular value, there are multiple other areas in the 1956 // system which believe this to be true (e.g. oop->object_size in some 1957 // cases incorrectly returns the size in wordSize units rather than 1958 // HeapWordSize). 1959 guarantee(HeapWordSize == wordSize, "HeapWordSize must equal wordSize"); 1960 1961 size_t init_byte_size = collector_policy()->initial_heap_byte_size(); 1962 size_t max_byte_size = collector_policy()->max_heap_byte_size(); 1963 1964 // Ensure that the sizes are properly aligned. 1965 Universe::check_alignment(init_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1966 Universe::check_alignment(max_byte_size, HeapRegion::GrainBytes, "g1 heap"); 1967 1968 _cg1r = new ConcurrentG1Refine(); 1969 1970 // Reserve the maximum. 1971 PermanentGenerationSpec* pgs = collector_policy()->permanent_generation(); 1972 // Includes the perm-gen. 1973 1974 // When compressed oops are enabled, the preferred heap base 1975 // is calculated by subtracting the requested size from the 1976 // 32Gb boundary and using the result as the base address for 1977 // heap reservation. If the requested size is not aligned to 1978 // HeapRegion::GrainBytes (i.e. the alignment that is passed 1979 // into the ReservedHeapSpace constructor) then the actual 1980 // base of the reserved heap may end up differing from the 1981 // address that was requested (i.e. the preferred heap base). 1982 // If this happens then we could end up using a non-optimal 1983 // compressed oops mode. 1984 1985 // Since max_byte_size is aligned to the size of a heap region (checked 1986 // above), we also need to align the perm gen size as it might not be. 1987 const size_t total_reserved = max_byte_size + 1988 align_size_up(pgs->max_size(), HeapRegion::GrainBytes); 1989 Universe::check_alignment(total_reserved, HeapRegion::GrainBytes, "g1 heap and perm"); 1990 1991 char* addr = Universe::preferred_heap_base(total_reserved, Universe::UnscaledNarrowOop); 1992 1993 ReservedHeapSpace heap_rs(total_reserved, HeapRegion::GrainBytes, 1994 UseLargePages, addr); 1995 1996 if (UseCompressedOops) { 1997 if (addr != NULL && !heap_rs.is_reserved()) { 1998 // Failed to reserve at specified address - the requested memory 1999 // region is taken already, for example, by 'java' launcher. 2000 // Try again to reserver heap higher. 2001 addr = Universe::preferred_heap_base(total_reserved, Universe::ZeroBasedNarrowOop); 2002 2003 ReservedHeapSpace heap_rs0(total_reserved, HeapRegion::GrainBytes, 2004 UseLargePages, addr); 2005 2006 if (addr != NULL && !heap_rs0.is_reserved()) { 2007 // Failed to reserve at specified address again - give up. 2008 addr = Universe::preferred_heap_base(total_reserved, Universe::HeapBasedNarrowOop); 2009 assert(addr == NULL, ""); 2010 2011 ReservedHeapSpace heap_rs1(total_reserved, HeapRegion::GrainBytes, 2012 UseLargePages, addr); 2013 heap_rs = heap_rs1; 2014 } else { 2015 heap_rs = heap_rs0; 2016 } 2017 } 2018 } 2019 2020 if (!heap_rs.is_reserved()) { 2021 vm_exit_during_initialization("Could not reserve enough space for object heap"); 2022 return JNI_ENOMEM; 2023 } 2024 2025 // It is important to do this in a way such that concurrent readers can't 2026 // temporarily think somethings in the heap. (I've actually seen this 2027 // happen in asserts: DLD.) 2028 _reserved.set_word_size(0); 2029 _reserved.set_start((HeapWord*)heap_rs.base()); 2030 _reserved.set_end((HeapWord*)(heap_rs.base() + heap_rs.size())); 2031 2032 _expansion_regions = (uint) (max_byte_size / HeapRegion::GrainBytes); 2033 2034 // Create the gen rem set (and barrier set) for the entire reserved region. 2035 _rem_set = collector_policy()->create_rem_set(_reserved, 2); 2036 set_barrier_set(rem_set()->bs()); 2037 if (barrier_set()->is_a(BarrierSet::ModRef)) { 2038 _mr_bs = (ModRefBarrierSet*)_barrier_set; 2039 } else { 2040 vm_exit_during_initialization("G1 requires a mod ref bs."); 2041 return JNI_ENOMEM; 2042 } 2043 2044 // Also create a G1 rem set. 2045 if (mr_bs()->is_a(BarrierSet::CardTableModRef)) { 2046 _g1_rem_set = new G1RemSet(this, (CardTableModRefBS*)mr_bs()); 2047 } else { 2048 vm_exit_during_initialization("G1 requires a cardtable mod ref bs."); 2049 return JNI_ENOMEM; 2050 } 2051 2052 // Carve out the G1 part of the heap. 2053 2054 ReservedSpace g1_rs = heap_rs.first_part(max_byte_size); 2055 _g1_reserved = MemRegion((HeapWord*)g1_rs.base(), 2056 g1_rs.size()/HeapWordSize); 2057 ReservedSpace perm_gen_rs = heap_rs.last_part(max_byte_size); 2058 2059 _perm_gen = pgs->init(perm_gen_rs, pgs->init_size(), rem_set()); 2060 2061 _g1_storage.initialize(g1_rs, 0); 2062 _g1_committed = MemRegion((HeapWord*)_g1_storage.low(), (size_t) 0); 2063 _hrs.initialize((HeapWord*) _g1_reserved.start(), 2064 (HeapWord*) _g1_reserved.end(), 2065 _expansion_regions); 2066 2067 // 6843694 - ensure that the maximum region index can fit 2068 // in the remembered set structures. 2069 const uint max_region_idx = (1U << (sizeof(RegionIdx_t)*BitsPerByte-1)) - 1; 2070 guarantee((max_regions() - 1) <= max_region_idx, "too many regions"); 2071 2072 size_t max_cards_per_region = ((size_t)1 << (sizeof(CardIdx_t)*BitsPerByte-1)) - 1; 2073 guarantee(HeapRegion::CardsPerRegion > 0, "make sure it's initialized"); 2074 guarantee(HeapRegion::CardsPerRegion < max_cards_per_region, 2075 "too many cards per region"); 2076 2077 HeapRegionSet::set_unrealistically_long_length(max_regions() + 1); 2078 2079 _bot_shared = new G1BlockOffsetSharedArray(_reserved, 2080 heap_word_size(init_byte_size)); 2081 2082 _g1h = this; 2083 2084 _in_cset_fast_test_length = max_regions(); 2085 _in_cset_fast_test_base = 2086 NEW_C_HEAP_ARRAY(bool, (size_t) _in_cset_fast_test_length, mtGC); 2087 2088 // We're biasing _in_cset_fast_test to avoid subtracting the 2089 // beginning of the heap every time we want to index; basically 2090 // it's the same with what we do with the card table. 2091 _in_cset_fast_test = _in_cset_fast_test_base - 2092 ((uintx) _g1_reserved.start() >> HeapRegion::LogOfHRGrainBytes); 2093 2094 // Clear the _cset_fast_test bitmap in anticipation of adding 2095 // regions to the incremental collection set for the first 2096 // evacuation pause. 2097 clear_cset_fast_test(); 2098 2099 // Create the ConcurrentMark data structure and thread. 2100 // (Must do this late, so that "max_regions" is defined.) 2101 _cm = new ConcurrentMark(heap_rs, max_regions()); 2102 _cmThread = _cm->cmThread(); 2103 2104 // Initialize the from_card cache structure of HeapRegionRemSet. 2105 HeapRegionRemSet::init_heap(max_regions()); 2106 2107 // Now expand into the initial heap size. 2108 if (!expand(init_byte_size)) { 2109 vm_exit_during_initialization("Failed to allocate initial heap."); 2110 return JNI_ENOMEM; 2111 } 2112 2113 // Perform any initialization actions delegated to the policy. 2114 g1_policy()->init(); 2115 2116 _refine_cte_cl = 2117 new RefineCardTableEntryClosure(ConcurrentG1RefineThread::sts(), 2118 g1_rem_set(), 2119 concurrent_g1_refine()); 2120 JavaThread::dirty_card_queue_set().set_closure(_refine_cte_cl); 2121 2122 JavaThread::satb_mark_queue_set().initialize(SATB_Q_CBL_mon, 2123 SATB_Q_FL_lock, 2124 G1SATBProcessCompletedThreshold, 2125 Shared_SATB_Q_lock); 2126 2127 JavaThread::dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2128 DirtyCardQ_FL_lock, 2129 concurrent_g1_refine()->yellow_zone(), 2130 concurrent_g1_refine()->red_zone(), 2131 Shared_DirtyCardQ_lock); 2132 2133 if (G1DeferredRSUpdate) { 2134 dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon, 2135 DirtyCardQ_FL_lock, 2136 -1, // never trigger processing 2137 -1, // no limit on length 2138 Shared_DirtyCardQ_lock, 2139 &JavaThread::dirty_card_queue_set()); 2140 } 2141 2142 // Initialize the card queue set used to hold cards containing 2143 // references into the collection set. 2144 _into_cset_dirty_card_queue_set.initialize(DirtyCardQ_CBL_mon, 2145 DirtyCardQ_FL_lock, 2146 -1, // never trigger processing 2147 -1, // no limit on length 2148 Shared_DirtyCardQ_lock, 2149 &JavaThread::dirty_card_queue_set()); 2150 2151 // In case we're keeping closure specialization stats, initialize those 2152 // counts and that mechanism. 2153 SpecializationStats::clear(); 2154 2155 // Do later initialization work for concurrent refinement. 2156 _cg1r->init(); 2157 2158 // Here we allocate the dummy full region that is required by the 2159 // G1AllocRegion class. If we don't pass an address in the reserved 2160 // space here, lots of asserts fire. 2161 2162 HeapRegion* dummy_region = new_heap_region(0 /* index of bottom region */, 2163 _g1_reserved.start()); 2164 // We'll re-use the same region whether the alloc region will 2165 // require BOT updates or not and, if it doesn't, then a non-young 2166 // region will complain that it cannot support allocations without 2167 // BOT updates. So we'll tag the dummy region as young to avoid that. 2168 dummy_region->set_young(); 2169 // Make sure it's full. 2170 dummy_region->set_top(dummy_region->end()); 2171 G1AllocRegion::setup(this, dummy_region); 2172 2173 init_mutator_alloc_region(); 2174 2175 // Do create of the monitoring and management support so that 2176 // values in the heap have been properly initialized. 2177 _g1mm = new G1MonitoringSupport(this); 2178 2179 return JNI_OK; 2180 } 2181 2182 void G1CollectedHeap::ref_processing_init() { 2183 // Reference processing in G1 currently works as follows: 2184 // 2185 // * There are two reference processor instances. One is 2186 // used to record and process discovered references 2187 // during concurrent marking; the other is used to 2188 // record and process references during STW pauses 2189 // (both full and incremental). 2190 // * Both ref processors need to 'span' the entire heap as 2191 // the regions in the collection set may be dotted around. 2192 // 2193 // * For the concurrent marking ref processor: 2194 // * Reference discovery is enabled at initial marking. 2195 // * Reference discovery is disabled and the discovered 2196 // references processed etc during remarking. 2197 // * Reference discovery is MT (see below). 2198 // * Reference discovery requires a barrier (see below). 2199 // * Reference processing may or may not be MT 2200 // (depending on the value of ParallelRefProcEnabled 2201 // and ParallelGCThreads). 2202 // * A full GC disables reference discovery by the CM 2203 // ref processor and abandons any entries on it's 2204 // discovered lists. 2205 // 2206 // * For the STW processor: 2207 // * Non MT discovery is enabled at the start of a full GC. 2208 // * Processing and enqueueing during a full GC is non-MT. 2209 // * During a full GC, references are processed after marking. 2210 // 2211 // * Discovery (may or may not be MT) is enabled at the start 2212 // of an incremental evacuation pause. 2213 // * References are processed near the end of a STW evacuation pause. 2214 // * For both types of GC: 2215 // * Discovery is atomic - i.e. not concurrent. 2216 // * Reference discovery will not need a barrier. 2217 2218 SharedHeap::ref_processing_init(); 2219 MemRegion mr = reserved_region(); 2220 2221 // Concurrent Mark ref processor 2222 _ref_processor_cm = 2223 new ReferenceProcessor(mr, // span 2224 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2225 // mt processing 2226 (int) ParallelGCThreads, 2227 // degree of mt processing 2228 (ParallelGCThreads > 1) || (ConcGCThreads > 1), 2229 // mt discovery 2230 (int) MAX2(ParallelGCThreads, ConcGCThreads), 2231 // degree of mt discovery 2232 false, 2233 // Reference discovery is not atomic 2234 &_is_alive_closure_cm, 2235 // is alive closure 2236 // (for efficiency/performance) 2237 true); 2238 // Setting next fields of discovered 2239 // lists requires a barrier. 2240 2241 // STW ref processor 2242 _ref_processor_stw = 2243 new ReferenceProcessor(mr, // span 2244 ParallelRefProcEnabled && (ParallelGCThreads > 1), 2245 // mt processing 2246 MAX2((int)ParallelGCThreads, 1), 2247 // degree of mt processing 2248 (ParallelGCThreads > 1), 2249 // mt discovery 2250 MAX2((int)ParallelGCThreads, 1), 2251 // degree of mt discovery 2252 true, 2253 // Reference discovery is atomic 2254 &_is_alive_closure_stw, 2255 // is alive closure 2256 // (for efficiency/performance) 2257 false); 2258 // Setting next fields of discovered 2259 // lists requires a barrier. 2260 } 2261 2262 size_t G1CollectedHeap::capacity() const { 2263 return _g1_committed.byte_size(); 2264 } 2265 2266 void G1CollectedHeap::iterate_dirty_card_closure(CardTableEntryClosure* cl, 2267 DirtyCardQueue* into_cset_dcq, 2268 bool concurrent, 2269 int worker_i) { 2270 // Clean cards in the hot card cache 2271 concurrent_g1_refine()->clean_up_cache(worker_i, g1_rem_set(), into_cset_dcq); 2272 2273 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 2274 int n_completed_buffers = 0; 2275 while (dcqs.apply_closure_to_completed_buffer(cl, worker_i, 0, true)) { 2276 n_completed_buffers++; 2277 } 2278 g1_policy()->phase_times()->record_update_rs_processed_buffers(worker_i, n_completed_buffers); 2279 dcqs.clear_n_completed_buffers(); 2280 assert(!dcqs.completed_buffers_exist_dirty(), "Completed buffers exist!"); 2281 } 2282 2283 2284 // Computes the sum of the storage used by the various regions. 2285 2286 size_t G1CollectedHeap::used() const { 2287 assert(Heap_lock->owner() != NULL, 2288 "Should be owned on this thread's behalf."); 2289 size_t result = _summary_bytes_used; 2290 // Read only once in case it is set to NULL concurrently 2291 HeapRegion* hr = _mutator_alloc_region.get(); 2292 if (hr != NULL) 2293 result += hr->used(); 2294 return result; 2295 } 2296 2297 size_t G1CollectedHeap::used_unlocked() const { 2298 size_t result = _summary_bytes_used; 2299 return result; 2300 } 2301 2302 class SumUsedClosure: public HeapRegionClosure { 2303 size_t _used; 2304 public: 2305 SumUsedClosure() : _used(0) {} 2306 bool doHeapRegion(HeapRegion* r) { 2307 if (!r->continuesHumongous()) { 2308 _used += r->used(); 2309 } 2310 return false; 2311 } 2312 size_t result() { return _used; } 2313 }; 2314 2315 size_t G1CollectedHeap::recalculate_used() const { 2316 SumUsedClosure blk; 2317 heap_region_iterate(&blk); 2318 return blk.result(); 2319 } 2320 2321 size_t G1CollectedHeap::unsafe_max_alloc() { 2322 if (free_regions() > 0) return HeapRegion::GrainBytes; 2323 // otherwise, is there space in the current allocation region? 2324 2325 // We need to store the current allocation region in a local variable 2326 // here. The problem is that this method doesn't take any locks and 2327 // there may be other threads which overwrite the current allocation 2328 // region field. attempt_allocation(), for example, sets it to NULL 2329 // and this can happen *after* the NULL check here but before the call 2330 // to free(), resulting in a SIGSEGV. Note that this doesn't appear 2331 // to be a problem in the optimized build, since the two loads of the 2332 // current allocation region field are optimized away. 2333 HeapRegion* hr = _mutator_alloc_region.get(); 2334 if (hr == NULL) { 2335 return 0; 2336 } 2337 return hr->free(); 2338 } 2339 2340 bool G1CollectedHeap::should_do_concurrent_full_gc(GCCause::Cause cause) { 2341 switch (cause) { 2342 case GCCause::_gc_locker: return GCLockerInvokesConcurrent; 2343 case GCCause::_java_lang_system_gc: return ExplicitGCInvokesConcurrent; 2344 case GCCause::_g1_humongous_allocation: return true; 2345 default: return false; 2346 } 2347 } 2348 2349 #ifndef PRODUCT 2350 void G1CollectedHeap::allocate_dummy_regions() { 2351 // Let's fill up most of the region 2352 size_t word_size = HeapRegion::GrainWords - 1024; 2353 // And as a result the region we'll allocate will be humongous. 2354 guarantee(isHumongous(word_size), "sanity"); 2355 2356 for (uintx i = 0; i < G1DummyRegionsPerGC; ++i) { 2357 // Let's use the existing mechanism for the allocation 2358 HeapWord* dummy_obj = humongous_obj_allocate(word_size); 2359 if (dummy_obj != NULL) { 2360 MemRegion mr(dummy_obj, word_size); 2361 CollectedHeap::fill_with_object(mr); 2362 } else { 2363 // If we can't allocate once, we probably cannot allocate 2364 // again. Let's get out of the loop. 2365 break; 2366 } 2367 } 2368 } 2369 #endif // !PRODUCT 2370 2371 void G1CollectedHeap::increment_old_marking_cycles_started() { 2372 assert(_old_marking_cycles_started == _old_marking_cycles_completed || 2373 _old_marking_cycles_started == _old_marking_cycles_completed + 1, 2374 err_msg("Wrong marking cycle count (started: %d, completed: %d)", 2375 _old_marking_cycles_started, _old_marking_cycles_completed)); 2376 2377 _old_marking_cycles_started++; 2378 } 2379 2380 void G1CollectedHeap::increment_old_marking_cycles_completed(bool concurrent) { 2381 MonitorLockerEx x(FullGCCount_lock, Mutex::_no_safepoint_check_flag); 2382 2383 // We assume that if concurrent == true, then the caller is a 2384 // concurrent thread that was joined the Suspendible Thread 2385 // Set. If there's ever a cheap way to check this, we should add an 2386 // assert here. 2387 2388 // Given that this method is called at the end of a Full GC or of a 2389 // concurrent cycle, and those can be nested (i.e., a Full GC can 2390 // interrupt a concurrent cycle), the number of full collections 2391 // completed should be either one (in the case where there was no 2392 // nesting) or two (when a Full GC interrupted a concurrent cycle) 2393 // behind the number of full collections started. 2394 2395 // This is the case for the inner caller, i.e. a Full GC. 2396 assert(concurrent || 2397 (_old_marking_cycles_started == _old_marking_cycles_completed + 1) || 2398 (_old_marking_cycles_started == _old_marking_cycles_completed + 2), 2399 err_msg("for inner caller (Full GC): _old_marking_cycles_started = %u " 2400 "is inconsistent with _old_marking_cycles_completed = %u", 2401 _old_marking_cycles_started, _old_marking_cycles_completed)); 2402 2403 // This is the case for the outer caller, i.e. the concurrent cycle. 2404 assert(!concurrent || 2405 (_old_marking_cycles_started == _old_marking_cycles_completed + 1), 2406 err_msg("for outer caller (concurrent cycle): " 2407 "_old_marking_cycles_started = %u " 2408 "is inconsistent with _old_marking_cycles_completed = %u", 2409 _old_marking_cycles_started, _old_marking_cycles_completed)); 2410 2411 _old_marking_cycles_completed += 1; 2412 2413 // We need to clear the "in_progress" flag in the CM thread before 2414 // we wake up any waiters (especially when ExplicitInvokesConcurrent 2415 // is set) so that if a waiter requests another System.gc() it doesn't 2416 // incorrectly see that a marking cyle is still in progress. 2417 if (concurrent) { 2418 _cmThread->clear_in_progress(); 2419 } 2420 2421 // This notify_all() will ensure that a thread that called 2422 // System.gc() with (with ExplicitGCInvokesConcurrent set or not) 2423 // and it's waiting for a full GC to finish will be woken up. It is 2424 // waiting in VM_G1IncCollectionPause::doit_epilogue(). 2425 FullGCCount_lock->notify_all(); 2426 } 2427 2428 void G1CollectedHeap::collect_as_vm_thread(GCCause::Cause cause) { 2429 assert_at_safepoint(true /* should_be_vm_thread */); 2430 GCCauseSetter gcs(this, cause); 2431 switch (cause) { 2432 case GCCause::_heap_inspection: 2433 case GCCause::_heap_dump: { 2434 HandleMark hm; 2435 do_full_collection(false); // don't clear all soft refs 2436 break; 2437 } 2438 default: // XXX FIX ME 2439 ShouldNotReachHere(); // Unexpected use of this function 2440 } 2441 } 2442 2443 void G1CollectedHeap::collect(GCCause::Cause cause) { 2444 assert_heap_not_locked(); 2445 2446 unsigned int gc_count_before; 2447 unsigned int old_marking_count_before; 2448 bool retry_gc; 2449 2450 do { 2451 retry_gc = false; 2452 2453 { 2454 MutexLocker ml(Heap_lock); 2455 2456 // Read the GC count while holding the Heap_lock 2457 gc_count_before = total_collections(); 2458 old_marking_count_before = _old_marking_cycles_started; 2459 } 2460 2461 if (should_do_concurrent_full_gc(cause)) { 2462 // Schedule an initial-mark evacuation pause that will start a 2463 // concurrent cycle. We're setting word_size to 0 which means that 2464 // we are not requesting a post-GC allocation. 2465 VM_G1IncCollectionPause op(gc_count_before, 2466 0, /* word_size */ 2467 true, /* should_initiate_conc_mark */ 2468 g1_policy()->max_pause_time_ms(), 2469 cause); 2470 2471 VMThread::execute(&op); 2472 if (!op.pause_succeeded()) { 2473 if (old_marking_count_before == _old_marking_cycles_started) { 2474 retry_gc = op.should_retry_gc(); 2475 } else { 2476 // A Full GC happened while we were trying to schedule the 2477 // initial-mark GC. No point in starting a new cycle given 2478 // that the whole heap was collected anyway. 2479 } 2480 2481 if (retry_gc) { 2482 if (GC_locker::is_active_and_needs_gc()) { 2483 GC_locker::stall_until_clear(); 2484 } 2485 } 2486 } 2487 } else { 2488 if (cause == GCCause::_gc_locker 2489 DEBUG_ONLY(|| cause == GCCause::_scavenge_alot)) { 2490 2491 // Schedule a standard evacuation pause. We're setting word_size 2492 // to 0 which means that we are not requesting a post-GC allocation. 2493 VM_G1IncCollectionPause op(gc_count_before, 2494 0, /* word_size */ 2495 false, /* should_initiate_conc_mark */ 2496 g1_policy()->max_pause_time_ms(), 2497 cause); 2498 VMThread::execute(&op); 2499 } else { 2500 // Schedule a Full GC. 2501 VM_G1CollectFull op(gc_count_before, old_marking_count_before, cause); 2502 VMThread::execute(&op); 2503 } 2504 } 2505 } while (retry_gc); 2506 } 2507 2508 bool G1CollectedHeap::is_in(const void* p) const { 2509 if (_g1_committed.contains(p)) { 2510 // Given that we know that p is in the committed space, 2511 // heap_region_containing_raw() should successfully 2512 // return the containing region. 2513 HeapRegion* hr = heap_region_containing_raw(p); 2514 return hr->is_in(p); 2515 } else { 2516 return _perm_gen->as_gen()->is_in(p); 2517 } 2518 } 2519 2520 // Iteration functions. 2521 2522 // Iterates an OopClosure over all ref-containing fields of objects 2523 // within a HeapRegion. 2524 2525 class IterateOopClosureRegionClosure: public HeapRegionClosure { 2526 MemRegion _mr; 2527 OopClosure* _cl; 2528 public: 2529 IterateOopClosureRegionClosure(MemRegion mr, OopClosure* cl) 2530 : _mr(mr), _cl(cl) {} 2531 bool doHeapRegion(HeapRegion* r) { 2532 if (! r->continuesHumongous()) { 2533 r->oop_iterate(_cl); 2534 } 2535 return false; 2536 } 2537 }; 2538 2539 void G1CollectedHeap::oop_iterate(OopClosure* cl, bool do_perm) { 2540 IterateOopClosureRegionClosure blk(_g1_committed, cl); 2541 heap_region_iterate(&blk); 2542 if (do_perm) { 2543 perm_gen()->oop_iterate(cl); 2544 } 2545 } 2546 2547 void G1CollectedHeap::oop_iterate(MemRegion mr, OopClosure* cl, bool do_perm) { 2548 IterateOopClosureRegionClosure blk(mr, cl); 2549 heap_region_iterate(&blk); 2550 if (do_perm) { 2551 perm_gen()->oop_iterate(cl); 2552 } 2553 } 2554 2555 // Iterates an ObjectClosure over all objects within a HeapRegion. 2556 2557 class IterateObjectClosureRegionClosure: public HeapRegionClosure { 2558 ObjectClosure* _cl; 2559 public: 2560 IterateObjectClosureRegionClosure(ObjectClosure* cl) : _cl(cl) {} 2561 bool doHeapRegion(HeapRegion* r) { 2562 if (! r->continuesHumongous()) { 2563 r->object_iterate(_cl); 2564 } 2565 return false; 2566 } 2567 }; 2568 2569 void G1CollectedHeap::object_iterate(ObjectClosure* cl, bool do_perm) { 2570 IterateObjectClosureRegionClosure blk(cl); 2571 heap_region_iterate(&blk); 2572 if (do_perm) { 2573 perm_gen()->object_iterate(cl); 2574 } 2575 } 2576 2577 void G1CollectedHeap::object_iterate_since_last_GC(ObjectClosure* cl) { 2578 // FIXME: is this right? 2579 guarantee(false, "object_iterate_since_last_GC not supported by G1 heap"); 2580 } 2581 2582 // Calls a SpaceClosure on a HeapRegion. 2583 2584 class SpaceClosureRegionClosure: public HeapRegionClosure { 2585 SpaceClosure* _cl; 2586 public: 2587 SpaceClosureRegionClosure(SpaceClosure* cl) : _cl(cl) {} 2588 bool doHeapRegion(HeapRegion* r) { 2589 _cl->do_space(r); 2590 return false; 2591 } 2592 }; 2593 2594 void G1CollectedHeap::space_iterate(SpaceClosure* cl) { 2595 SpaceClosureRegionClosure blk(cl); 2596 heap_region_iterate(&blk); 2597 } 2598 2599 void G1CollectedHeap::heap_region_iterate(HeapRegionClosure* cl) const { 2600 _hrs.iterate(cl); 2601 } 2602 2603 void G1CollectedHeap::heap_region_iterate_from(HeapRegion* r, 2604 HeapRegionClosure* cl) const { 2605 _hrs.iterate_from(r, cl); 2606 } 2607 2608 void 2609 G1CollectedHeap::heap_region_par_iterate_chunked(HeapRegionClosure* cl, 2610 uint worker, 2611 uint no_of_par_workers, 2612 jint claim_value) { 2613 const uint regions = n_regions(); 2614 const uint max_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 2615 no_of_par_workers : 2616 1); 2617 assert(UseDynamicNumberOfGCThreads || 2618 no_of_par_workers == workers()->total_workers(), 2619 "Non dynamic should use fixed number of workers"); 2620 // try to spread out the starting points of the workers 2621 const uint start_index = regions / max_workers * worker; 2622 2623 // each worker will actually look at all regions 2624 for (uint count = 0; count < regions; ++count) { 2625 const uint index = (start_index + count) % regions; 2626 assert(0 <= index && index < regions, "sanity"); 2627 HeapRegion* r = region_at(index); 2628 // we'll ignore "continues humongous" regions (we'll process them 2629 // when we come across their corresponding "start humongous" 2630 // region) and regions already claimed 2631 if (r->claim_value() == claim_value || r->continuesHumongous()) { 2632 continue; 2633 } 2634 // OK, try to claim it 2635 if (r->claimHeapRegion(claim_value)) { 2636 // success! 2637 assert(!r->continuesHumongous(), "sanity"); 2638 if (r->startsHumongous()) { 2639 // If the region is "starts humongous" we'll iterate over its 2640 // "continues humongous" first; in fact we'll do them 2641 // first. The order is important. In on case, calling the 2642 // closure on the "starts humongous" region might de-allocate 2643 // and clear all its "continues humongous" regions and, as a 2644 // result, we might end up processing them twice. So, we'll do 2645 // them first (notice: most closures will ignore them anyway) and 2646 // then we'll do the "starts humongous" region. 2647 for (uint ch_index = index + 1; ch_index < regions; ++ch_index) { 2648 HeapRegion* chr = region_at(ch_index); 2649 2650 // if the region has already been claimed or it's not 2651 // "continues humongous" we're done 2652 if (chr->claim_value() == claim_value || 2653 !chr->continuesHumongous()) { 2654 break; 2655 } 2656 2657 // Noone should have claimed it directly. We can given 2658 // that we claimed its "starts humongous" region. 2659 assert(chr->claim_value() != claim_value, "sanity"); 2660 assert(chr->humongous_start_region() == r, "sanity"); 2661 2662 if (chr->claimHeapRegion(claim_value)) { 2663 // we should always be able to claim it; noone else should 2664 // be trying to claim this region 2665 2666 bool res2 = cl->doHeapRegion(chr); 2667 assert(!res2, "Should not abort"); 2668 2669 // Right now, this holds (i.e., no closure that actually 2670 // does something with "continues humongous" regions 2671 // clears them). We might have to weaken it in the future, 2672 // but let's leave these two asserts here for extra safety. 2673 assert(chr->continuesHumongous(), "should still be the case"); 2674 assert(chr->humongous_start_region() == r, "sanity"); 2675 } else { 2676 guarantee(false, "we should not reach here"); 2677 } 2678 } 2679 } 2680 2681 assert(!r->continuesHumongous(), "sanity"); 2682 bool res = cl->doHeapRegion(r); 2683 assert(!res, "Should not abort"); 2684 } 2685 } 2686 } 2687 2688 class ResetClaimValuesClosure: public HeapRegionClosure { 2689 public: 2690 bool doHeapRegion(HeapRegion* r) { 2691 r->set_claim_value(HeapRegion::InitialClaimValue); 2692 return false; 2693 } 2694 }; 2695 2696 void G1CollectedHeap::reset_heap_region_claim_values() { 2697 ResetClaimValuesClosure blk; 2698 heap_region_iterate(&blk); 2699 } 2700 2701 void G1CollectedHeap::reset_cset_heap_region_claim_values() { 2702 ResetClaimValuesClosure blk; 2703 collection_set_iterate(&blk); 2704 } 2705 2706 #ifdef ASSERT 2707 // This checks whether all regions in the heap have the correct claim 2708 // value. I also piggy-backed on this a check to ensure that the 2709 // humongous_start_region() information on "continues humongous" 2710 // regions is correct. 2711 2712 class CheckClaimValuesClosure : public HeapRegionClosure { 2713 private: 2714 jint _claim_value; 2715 uint _failures; 2716 HeapRegion* _sh_region; 2717 2718 public: 2719 CheckClaimValuesClosure(jint claim_value) : 2720 _claim_value(claim_value), _failures(0), _sh_region(NULL) { } 2721 bool doHeapRegion(HeapRegion* r) { 2722 if (r->claim_value() != _claim_value) { 2723 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2724 "claim value = %d, should be %d", 2725 HR_FORMAT_PARAMS(r), 2726 r->claim_value(), _claim_value); 2727 ++_failures; 2728 } 2729 if (!r->isHumongous()) { 2730 _sh_region = NULL; 2731 } else if (r->startsHumongous()) { 2732 _sh_region = r; 2733 } else if (r->continuesHumongous()) { 2734 if (r->humongous_start_region() != _sh_region) { 2735 gclog_or_tty->print_cr("Region " HR_FORMAT ", " 2736 "HS = "PTR_FORMAT", should be "PTR_FORMAT, 2737 HR_FORMAT_PARAMS(r), 2738 r->humongous_start_region(), 2739 _sh_region); 2740 ++_failures; 2741 } 2742 } 2743 return false; 2744 } 2745 uint failures() { return _failures; } 2746 }; 2747 2748 bool G1CollectedHeap::check_heap_region_claim_values(jint claim_value) { 2749 CheckClaimValuesClosure cl(claim_value); 2750 heap_region_iterate(&cl); 2751 return cl.failures() == 0; 2752 } 2753 2754 class CheckClaimValuesInCSetHRClosure: public HeapRegionClosure { 2755 private: 2756 jint _claim_value; 2757 uint _failures; 2758 2759 public: 2760 CheckClaimValuesInCSetHRClosure(jint claim_value) : 2761 _claim_value(claim_value), _failures(0) { } 2762 2763 uint failures() { return _failures; } 2764 2765 bool doHeapRegion(HeapRegion* hr) { 2766 assert(hr->in_collection_set(), "how?"); 2767 assert(!hr->isHumongous(), "H-region in CSet"); 2768 if (hr->claim_value() != _claim_value) { 2769 gclog_or_tty->print_cr("CSet Region " HR_FORMAT ", " 2770 "claim value = %d, should be %d", 2771 HR_FORMAT_PARAMS(hr), 2772 hr->claim_value(), _claim_value); 2773 _failures += 1; 2774 } 2775 return false; 2776 } 2777 }; 2778 2779 bool G1CollectedHeap::check_cset_heap_region_claim_values(jint claim_value) { 2780 CheckClaimValuesInCSetHRClosure cl(claim_value); 2781 collection_set_iterate(&cl); 2782 return cl.failures() == 0; 2783 } 2784 #endif // ASSERT 2785 2786 // Clear the cached CSet starting regions and (more importantly) 2787 // the time stamps. Called when we reset the GC time stamp. 2788 void G1CollectedHeap::clear_cset_start_regions() { 2789 assert(_worker_cset_start_region != NULL, "sanity"); 2790 assert(_worker_cset_start_region_time_stamp != NULL, "sanity"); 2791 2792 int n_queues = MAX2((int)ParallelGCThreads, 1); 2793 for (int i = 0; i < n_queues; i++) { 2794 _worker_cset_start_region[i] = NULL; 2795 _worker_cset_start_region_time_stamp[i] = 0; 2796 } 2797 } 2798 2799 // Given the id of a worker, obtain or calculate a suitable 2800 // starting region for iterating over the current collection set. 2801 HeapRegion* G1CollectedHeap::start_cset_region_for_worker(int worker_i) { 2802 assert(get_gc_time_stamp() > 0, "should have been updated by now"); 2803 2804 HeapRegion* result = NULL; 2805 unsigned gc_time_stamp = get_gc_time_stamp(); 2806 2807 if (_worker_cset_start_region_time_stamp[worker_i] == gc_time_stamp) { 2808 // Cached starting region for current worker was set 2809 // during the current pause - so it's valid. 2810 // Note: the cached starting heap region may be NULL 2811 // (when the collection set is empty). 2812 result = _worker_cset_start_region[worker_i]; 2813 assert(result == NULL || result->in_collection_set(), "sanity"); 2814 return result; 2815 } 2816 2817 // The cached entry was not valid so let's calculate 2818 // a suitable starting heap region for this worker. 2819 2820 // We want the parallel threads to start their collection 2821 // set iteration at different collection set regions to 2822 // avoid contention. 2823 // If we have: 2824 // n collection set regions 2825 // p threads 2826 // Then thread t will start at region floor ((t * n) / p) 2827 2828 result = g1_policy()->collection_set(); 2829 if (G1CollectedHeap::use_parallel_gc_threads()) { 2830 uint cs_size = g1_policy()->cset_region_length(); 2831 uint active_workers = workers()->active_workers(); 2832 assert(UseDynamicNumberOfGCThreads || 2833 active_workers == workers()->total_workers(), 2834 "Unless dynamic should use total workers"); 2835 2836 uint end_ind = (cs_size * worker_i) / active_workers; 2837 uint start_ind = 0; 2838 2839 if (worker_i > 0 && 2840 _worker_cset_start_region_time_stamp[worker_i - 1] == gc_time_stamp) { 2841 // Previous workers starting region is valid 2842 // so let's iterate from there 2843 start_ind = (cs_size * (worker_i - 1)) / active_workers; 2844 result = _worker_cset_start_region[worker_i - 1]; 2845 } 2846 2847 for (uint i = start_ind; i < end_ind; i++) { 2848 result = result->next_in_collection_set(); 2849 } 2850 } 2851 2852 // Note: the calculated starting heap region may be NULL 2853 // (when the collection set is empty). 2854 assert(result == NULL || result->in_collection_set(), "sanity"); 2855 assert(_worker_cset_start_region_time_stamp[worker_i] != gc_time_stamp, 2856 "should be updated only once per pause"); 2857 _worker_cset_start_region[worker_i] = result; 2858 OrderAccess::storestore(); 2859 _worker_cset_start_region_time_stamp[worker_i] = gc_time_stamp; 2860 return result; 2861 } 2862 2863 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) { 2864 HeapRegion* r = g1_policy()->collection_set(); 2865 while (r != NULL) { 2866 HeapRegion* next = r->next_in_collection_set(); 2867 if (cl->doHeapRegion(r)) { 2868 cl->incomplete(); 2869 return; 2870 } 2871 r = next; 2872 } 2873 } 2874 2875 void G1CollectedHeap::collection_set_iterate_from(HeapRegion* r, 2876 HeapRegionClosure *cl) { 2877 if (r == NULL) { 2878 // The CSet is empty so there's nothing to do. 2879 return; 2880 } 2881 2882 assert(r->in_collection_set(), 2883 "Start region must be a member of the collection set."); 2884 HeapRegion* cur = r; 2885 while (cur != NULL) { 2886 HeapRegion* next = cur->next_in_collection_set(); 2887 if (cl->doHeapRegion(cur) && false) { 2888 cl->incomplete(); 2889 return; 2890 } 2891 cur = next; 2892 } 2893 cur = g1_policy()->collection_set(); 2894 while (cur != r) { 2895 HeapRegion* next = cur->next_in_collection_set(); 2896 if (cl->doHeapRegion(cur) && false) { 2897 cl->incomplete(); 2898 return; 2899 } 2900 cur = next; 2901 } 2902 } 2903 2904 CompactibleSpace* G1CollectedHeap::first_compactible_space() { 2905 return n_regions() > 0 ? region_at(0) : NULL; 2906 } 2907 2908 2909 Space* G1CollectedHeap::space_containing(const void* addr) const { 2910 Space* res = heap_region_containing(addr); 2911 if (res == NULL) 2912 res = perm_gen()->space_containing(addr); 2913 return res; 2914 } 2915 2916 HeapWord* G1CollectedHeap::block_start(const void* addr) const { 2917 Space* sp = space_containing(addr); 2918 if (sp != NULL) { 2919 return sp->block_start(addr); 2920 } 2921 return NULL; 2922 } 2923 2924 size_t G1CollectedHeap::block_size(const HeapWord* addr) const { 2925 Space* sp = space_containing(addr); 2926 assert(sp != NULL, "block_size of address outside of heap"); 2927 return sp->block_size(addr); 2928 } 2929 2930 bool G1CollectedHeap::block_is_obj(const HeapWord* addr) const { 2931 Space* sp = space_containing(addr); 2932 return sp->block_is_obj(addr); 2933 } 2934 2935 bool G1CollectedHeap::supports_tlab_allocation() const { 2936 return true; 2937 } 2938 2939 size_t G1CollectedHeap::tlab_capacity(Thread* ignored) const { 2940 return HeapRegion::GrainBytes; 2941 } 2942 2943 size_t G1CollectedHeap::unsafe_max_tlab_alloc(Thread* ignored) const { 2944 // Return the remaining space in the cur alloc region, but not less than 2945 // the min TLAB size. 2946 2947 // Also, this value can be at most the humongous object threshold, 2948 // since we can't allow tlabs to grow big enough to accomodate 2949 // humongous objects. 2950 2951 HeapRegion* hr = _mutator_alloc_region.get(); 2952 size_t max_tlab_size = _humongous_object_threshold_in_words * wordSize; 2953 if (hr == NULL) { 2954 return max_tlab_size; 2955 } else { 2956 return MIN2(MAX2(hr->free(), (size_t) MinTLABSize), max_tlab_size); 2957 } 2958 } 2959 2960 size_t G1CollectedHeap::max_capacity() const { 2961 return _g1_reserved.byte_size(); 2962 } 2963 2964 jlong G1CollectedHeap::millis_since_last_gc() { 2965 // assert(false, "NYI"); 2966 return 0; 2967 } 2968 2969 void G1CollectedHeap::prepare_for_verify() { 2970 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 2971 ensure_parsability(false); 2972 } 2973 g1_rem_set()->prepare_for_verify(); 2974 } 2975 2976 class VerifyLivenessOopClosure: public OopClosure { 2977 G1CollectedHeap* _g1h; 2978 VerifyOption _vo; 2979 public: 2980 VerifyLivenessOopClosure(G1CollectedHeap* g1h, VerifyOption vo): 2981 _g1h(g1h), _vo(vo) 2982 { } 2983 void do_oop(narrowOop *p) { do_oop_work(p); } 2984 void do_oop( oop *p) { do_oop_work(p); } 2985 2986 template <class T> void do_oop_work(T *p) { 2987 oop obj = oopDesc::load_decode_heap_oop(p); 2988 guarantee(obj == NULL || !_g1h->is_obj_dead_cond(obj, _vo), 2989 "Dead object referenced by a not dead object"); 2990 } 2991 }; 2992 2993 class VerifyObjsInRegionClosure: public ObjectClosure { 2994 private: 2995 G1CollectedHeap* _g1h; 2996 size_t _live_bytes; 2997 HeapRegion *_hr; 2998 VerifyOption _vo; 2999 public: 3000 // _vo == UsePrevMarking -> use "prev" marking information, 3001 // _vo == UseNextMarking -> use "next" marking information, 3002 // _vo == UseMarkWord -> use mark word from object header. 3003 VerifyObjsInRegionClosure(HeapRegion *hr, VerifyOption vo) 3004 : _live_bytes(0), _hr(hr), _vo(vo) { 3005 _g1h = G1CollectedHeap::heap(); 3006 } 3007 void do_object(oop o) { 3008 VerifyLivenessOopClosure isLive(_g1h, _vo); 3009 assert(o != NULL, "Huh?"); 3010 if (!_g1h->is_obj_dead_cond(o, _vo)) { 3011 // If the object is alive according to the mark word, 3012 // then verify that the marking information agrees. 3013 // Note we can't verify the contra-positive of the 3014 // above: if the object is dead (according to the mark 3015 // word), it may not be marked, or may have been marked 3016 // but has since became dead, or may have been allocated 3017 // since the last marking. 3018 if (_vo == VerifyOption_G1UseMarkWord) { 3019 guarantee(!_g1h->is_obj_dead(o), "mark word and concurrent mark mismatch"); 3020 } 3021 3022 o->oop_iterate(&isLive); 3023 if (!_hr->obj_allocated_since_prev_marking(o)) { 3024 size_t obj_size = o->size(); // Make sure we don't overflow 3025 _live_bytes += (obj_size * HeapWordSize); 3026 } 3027 } 3028 } 3029 size_t live_bytes() { return _live_bytes; } 3030 }; 3031 3032 class PrintObjsInRegionClosure : public ObjectClosure { 3033 HeapRegion *_hr; 3034 G1CollectedHeap *_g1; 3035 public: 3036 PrintObjsInRegionClosure(HeapRegion *hr) : _hr(hr) { 3037 _g1 = G1CollectedHeap::heap(); 3038 }; 3039 3040 void do_object(oop o) { 3041 if (o != NULL) { 3042 HeapWord *start = (HeapWord *) o; 3043 size_t word_sz = o->size(); 3044 gclog_or_tty->print("\nPrinting obj "PTR_FORMAT" of size " SIZE_FORMAT 3045 " isMarkedPrev %d isMarkedNext %d isAllocSince %d\n", 3046 (void*) o, word_sz, 3047 _g1->isMarkedPrev(o), 3048 _g1->isMarkedNext(o), 3049 _hr->obj_allocated_since_prev_marking(o)); 3050 HeapWord *end = start + word_sz; 3051 HeapWord *cur; 3052 int *val; 3053 for (cur = start; cur < end; cur++) { 3054 val = (int *) cur; 3055 gclog_or_tty->print("\t "PTR_FORMAT":"PTR_FORMAT"\n", val, *val); 3056 } 3057 } 3058 } 3059 }; 3060 3061 class VerifyRegionClosure: public HeapRegionClosure { 3062 private: 3063 bool _par; 3064 VerifyOption _vo; 3065 bool _failures; 3066 public: 3067 // _vo == UsePrevMarking -> use "prev" marking information, 3068 // _vo == UseNextMarking -> use "next" marking information, 3069 // _vo == UseMarkWord -> use mark word from object header. 3070 VerifyRegionClosure(bool par, VerifyOption vo) 3071 : _par(par), 3072 _vo(vo), 3073 _failures(false) {} 3074 3075 bool failures() { 3076 return _failures; 3077 } 3078 3079 bool doHeapRegion(HeapRegion* r) { 3080 guarantee(_par || r->claim_value() == HeapRegion::InitialClaimValue, 3081 "Should be unclaimed at verify points."); 3082 if (!r->continuesHumongous()) { 3083 bool failures = false; 3084 r->verify(_vo, &failures); 3085 if (failures) { 3086 _failures = true; 3087 } else { 3088 VerifyObjsInRegionClosure not_dead_yet_cl(r, _vo); 3089 r->object_iterate(¬_dead_yet_cl); 3090 if (_vo != VerifyOption_G1UseNextMarking) { 3091 if (r->max_live_bytes() < not_dead_yet_cl.live_bytes()) { 3092 gclog_or_tty->print_cr("["PTR_FORMAT","PTR_FORMAT"] " 3093 "max_live_bytes "SIZE_FORMAT" " 3094 "< calculated "SIZE_FORMAT, 3095 r->bottom(), r->end(), 3096 r->max_live_bytes(), 3097 not_dead_yet_cl.live_bytes()); 3098 _failures = true; 3099 } 3100 } else { 3101 // When vo == UseNextMarking we cannot currently do a sanity 3102 // check on the live bytes as the calculation has not been 3103 // finalized yet. 3104 } 3105 } 3106 } 3107 return false; // stop the region iteration if we hit a failure 3108 } 3109 }; 3110 3111 class VerifyRootsClosure: public OopsInGenClosure { 3112 private: 3113 G1CollectedHeap* _g1h; 3114 VerifyOption _vo; 3115 bool _failures; 3116 public: 3117 // _vo == UsePrevMarking -> use "prev" marking information, 3118 // _vo == UseNextMarking -> use "next" marking information, 3119 // _vo == UseMarkWord -> use mark word from object header. 3120 VerifyRootsClosure(VerifyOption vo) : 3121 _g1h(G1CollectedHeap::heap()), 3122 _vo(vo), 3123 _failures(false) { } 3124 3125 bool failures() { return _failures; } 3126 3127 template <class T> void do_oop_nv(T* p) { 3128 T heap_oop = oopDesc::load_heap_oop(p); 3129 if (!oopDesc::is_null(heap_oop)) { 3130 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); 3131 if (_g1h->is_obj_dead_cond(obj, _vo)) { 3132 gclog_or_tty->print_cr("Root location "PTR_FORMAT" " 3133 "points to dead obj "PTR_FORMAT, p, (void*) obj); 3134 if (_vo == VerifyOption_G1UseMarkWord) { 3135 gclog_or_tty->print_cr(" Mark word: "PTR_FORMAT, (void*)(obj->mark())); 3136 } 3137 obj->print_on(gclog_or_tty); 3138 _failures = true; 3139 } 3140 } 3141 } 3142 3143 void do_oop(oop* p) { do_oop_nv(p); } 3144 void do_oop(narrowOop* p) { do_oop_nv(p); } 3145 }; 3146 3147 // This is the task used for parallel heap verification. 3148 3149 class G1ParVerifyTask: public AbstractGangTask { 3150 private: 3151 G1CollectedHeap* _g1h; 3152 VerifyOption _vo; 3153 bool _failures; 3154 3155 public: 3156 // _vo == UsePrevMarking -> use "prev" marking information, 3157 // _vo == UseNextMarking -> use "next" marking information, 3158 // _vo == UseMarkWord -> use mark word from object header. 3159 G1ParVerifyTask(G1CollectedHeap* g1h, VerifyOption vo) : 3160 AbstractGangTask("Parallel verify task"), 3161 _g1h(g1h), 3162 _vo(vo), 3163 _failures(false) { } 3164 3165 bool failures() { 3166 return _failures; 3167 } 3168 3169 void work(uint worker_id) { 3170 HandleMark hm; 3171 VerifyRegionClosure blk(true, _vo); 3172 _g1h->heap_region_par_iterate_chunked(&blk, worker_id, 3173 _g1h->workers()->active_workers(), 3174 HeapRegion::ParVerifyClaimValue); 3175 if (blk.failures()) { 3176 _failures = true; 3177 } 3178 } 3179 }; 3180 3181 void G1CollectedHeap::verify(bool silent) { 3182 verify(silent, VerifyOption_G1UsePrevMarking); 3183 } 3184 3185 void G1CollectedHeap::verify(bool silent, 3186 VerifyOption vo) { 3187 if (SafepointSynchronize::is_at_safepoint() || ! UseTLAB) { 3188 if (!silent) { gclog_or_tty->print("Roots (excluding permgen) "); } 3189 VerifyRootsClosure rootsCl(vo); 3190 3191 assert(Thread::current()->is_VM_thread(), 3192 "Expected to be executed serially by the VM thread at this point"); 3193 3194 CodeBlobToOopClosure blobsCl(&rootsCl, /*do_marking=*/ false); 3195 3196 // We apply the relevant closures to all the oops in the 3197 // system dictionary, the string table and the code cache. 3198 const int so = SO_AllClasses | SO_Strings | SO_CodeCache; 3199 3200 process_strong_roots(true, // activate StrongRootsScope 3201 true, // we set "collecting perm gen" to true, 3202 // so we don't reset the dirty cards in the perm gen. 3203 ScanningOption(so), // roots scanning options 3204 &rootsCl, 3205 &blobsCl, 3206 &rootsCl); 3207 3208 // If we're verifying after the marking phase of a Full GC then we can't 3209 // treat the perm gen as roots into the G1 heap. Some of the objects in 3210 // the perm gen may be dead and hence not marked. If one of these dead 3211 // objects is considered to be a root then we may end up with a false 3212 // "Root location <x> points to dead ob <y>" failure. 3213 if (vo != VerifyOption_G1UseMarkWord) { 3214 // Since we used "collecting_perm_gen" == true above, we will not have 3215 // checked the refs from perm into the G1-collected heap. We check those 3216 // references explicitly below. Whether the relevant cards are dirty 3217 // is checked further below in the rem set verification. 3218 if (!silent) { gclog_or_tty->print("Permgen roots "); } 3219 perm_gen()->oop_iterate(&rootsCl); 3220 } 3221 bool failures = rootsCl.failures(); 3222 3223 if (vo != VerifyOption_G1UseMarkWord) { 3224 // If we're verifying during a full GC then the region sets 3225 // will have been torn down at the start of the GC. Therefore 3226 // verifying the region sets will fail. So we only verify 3227 // the region sets when not in a full GC. 3228 if (!silent) { gclog_or_tty->print("HeapRegionSets "); } 3229 verify_region_sets(); 3230 } 3231 3232 if (!silent) { gclog_or_tty->print("HeapRegions "); } 3233 if (GCParallelVerificationEnabled && ParallelGCThreads > 1) { 3234 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3235 "sanity check"); 3236 3237 G1ParVerifyTask task(this, vo); 3238 assert(UseDynamicNumberOfGCThreads || 3239 workers()->active_workers() == workers()->total_workers(), 3240 "If not dynamic should be using all the workers"); 3241 int n_workers = workers()->active_workers(); 3242 set_par_threads(n_workers); 3243 workers()->run_task(&task); 3244 set_par_threads(0); 3245 if (task.failures()) { 3246 failures = true; 3247 } 3248 3249 // Checks that the expected amount of parallel work was done. 3250 // The implication is that n_workers is > 0. 3251 assert(check_heap_region_claim_values(HeapRegion::ParVerifyClaimValue), 3252 "sanity check"); 3253 3254 reset_heap_region_claim_values(); 3255 3256 assert(check_heap_region_claim_values(HeapRegion::InitialClaimValue), 3257 "sanity check"); 3258 } else { 3259 VerifyRegionClosure blk(false, vo); 3260 heap_region_iterate(&blk); 3261 if (blk.failures()) { 3262 failures = true; 3263 } 3264 } 3265 if (!silent) gclog_or_tty->print("RemSet "); 3266 rem_set()->verify(); 3267 3268 if (failures) { 3269 gclog_or_tty->print_cr("Heap:"); 3270 // It helps to have the per-region information in the output to 3271 // help us track down what went wrong. This is why we call 3272 // print_extended_on() instead of print_on(). 3273 print_extended_on(gclog_or_tty); 3274 gclog_or_tty->print_cr(""); 3275 #ifndef PRODUCT 3276 if (VerifyDuringGC && G1VerifyDuringGCPrintReachable) { 3277 concurrent_mark()->print_reachable("at-verification-failure", 3278 vo, false /* all */); 3279 } 3280 #endif 3281 gclog_or_tty->flush(); 3282 } 3283 guarantee(!failures, "there should not have been any failures"); 3284 } else { 3285 if (!silent) gclog_or_tty->print("(SKIPPING roots, heapRegions, remset) "); 3286 } 3287 } 3288 3289 class PrintRegionClosure: public HeapRegionClosure { 3290 outputStream* _st; 3291 public: 3292 PrintRegionClosure(outputStream* st) : _st(st) {} 3293 bool doHeapRegion(HeapRegion* r) { 3294 r->print_on(_st); 3295 return false; 3296 } 3297 }; 3298 3299 void G1CollectedHeap::print_on(outputStream* st) const { 3300 st->print(" %-20s", "garbage-first heap"); 3301 st->print(" total " SIZE_FORMAT "K, used " SIZE_FORMAT "K", 3302 capacity()/K, used_unlocked()/K); 3303 st->print(" [" INTPTR_FORMAT ", " INTPTR_FORMAT ", " INTPTR_FORMAT ")", 3304 _g1_storage.low_boundary(), 3305 _g1_storage.high(), 3306 _g1_storage.high_boundary()); 3307 st->cr(); 3308 st->print(" region size " SIZE_FORMAT "K, ", HeapRegion::GrainBytes / K); 3309 uint young_regions = _young_list->length(); 3310 st->print("%u young (" SIZE_FORMAT "K), ", young_regions, 3311 (size_t) young_regions * HeapRegion::GrainBytes / K); 3312 uint survivor_regions = g1_policy()->recorded_survivor_regions(); 3313 st->print("%u survivors (" SIZE_FORMAT "K)", survivor_regions, 3314 (size_t) survivor_regions * HeapRegion::GrainBytes / K); 3315 st->cr(); 3316 perm()->as_gen()->print_on(st); 3317 } 3318 3319 void G1CollectedHeap::print_extended_on(outputStream* st) const { 3320 print_on(st); 3321 3322 // Print the per-region information. 3323 st->cr(); 3324 st->print_cr("Heap Regions: (Y=young(eden), SU=young(survivor), " 3325 "HS=humongous(starts), HC=humongous(continues), " 3326 "CS=collection set, F=free, TS=gc time stamp, " 3327 "PTAMS=previous top-at-mark-start, " 3328 "NTAMS=next top-at-mark-start)"); 3329 PrintRegionClosure blk(st); 3330 heap_region_iterate(&blk); 3331 } 3332 3333 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const { 3334 if (G1CollectedHeap::use_parallel_gc_threads()) { 3335 workers()->print_worker_threads_on(st); 3336 } 3337 _cmThread->print_on(st); 3338 st->cr(); 3339 _cm->print_worker_threads_on(st); 3340 _cg1r->print_worker_threads_on(st); 3341 st->cr(); 3342 } 3343 3344 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const { 3345 if (G1CollectedHeap::use_parallel_gc_threads()) { 3346 workers()->threads_do(tc); 3347 } 3348 tc->do_thread(_cmThread); 3349 _cg1r->threads_do(tc); 3350 } 3351 3352 void G1CollectedHeap::print_tracing_info() const { 3353 // We'll overload this to mean "trace GC pause statistics." 3354 if (TraceGen0Time || TraceGen1Time) { 3355 // The "G1CollectorPolicy" is keeping track of these stats, so delegate 3356 // to that. 3357 g1_policy()->print_tracing_info(); 3358 } 3359 if (G1SummarizeRSetStats) { 3360 g1_rem_set()->print_summary_info(); 3361 } 3362 if (G1SummarizeConcMark) { 3363 concurrent_mark()->print_summary_info(); 3364 } 3365 g1_policy()->print_yg_surv_rate_info(); 3366 SpecializationStats::print(); 3367 } 3368 3369 #ifndef PRODUCT 3370 // Helpful for debugging RSet issues. 3371 3372 class PrintRSetsClosure : public HeapRegionClosure { 3373 private: 3374 const char* _msg; 3375 size_t _occupied_sum; 3376 3377 public: 3378 bool doHeapRegion(HeapRegion* r) { 3379 HeapRegionRemSet* hrrs = r->rem_set(); 3380 size_t occupied = hrrs->occupied(); 3381 _occupied_sum += occupied; 3382 3383 gclog_or_tty->print_cr("Printing RSet for region "HR_FORMAT, 3384 HR_FORMAT_PARAMS(r)); 3385 if (occupied == 0) { 3386 gclog_or_tty->print_cr(" RSet is empty"); 3387 } else { 3388 hrrs->print(); 3389 } 3390 gclog_or_tty->print_cr("----------"); 3391 return false; 3392 } 3393 3394 PrintRSetsClosure(const char* msg) : _msg(msg), _occupied_sum(0) { 3395 gclog_or_tty->cr(); 3396 gclog_or_tty->print_cr("========================================"); 3397 gclog_or_tty->print_cr(msg); 3398 gclog_or_tty->cr(); 3399 } 3400 3401 ~PrintRSetsClosure() { 3402 gclog_or_tty->print_cr("Occupied Sum: "SIZE_FORMAT, _occupied_sum); 3403 gclog_or_tty->print_cr("========================================"); 3404 gclog_or_tty->cr(); 3405 } 3406 }; 3407 3408 void G1CollectedHeap::print_cset_rsets() { 3409 PrintRSetsClosure cl("Printing CSet RSets"); 3410 collection_set_iterate(&cl); 3411 } 3412 3413 void G1CollectedHeap::print_all_rsets() { 3414 PrintRSetsClosure cl("Printing All RSets");; 3415 heap_region_iterate(&cl); 3416 } 3417 #endif // PRODUCT 3418 3419 G1CollectedHeap* G1CollectedHeap::heap() { 3420 assert(_sh->kind() == CollectedHeap::G1CollectedHeap, 3421 "not a garbage-first heap"); 3422 return _g1h; 3423 } 3424 3425 void G1CollectedHeap::gc_prologue(bool full /* Ignored */) { 3426 // always_do_update_barrier = false; 3427 assert(InlineCacheBuffer::is_empty(), "should have cleaned up ICBuffer"); 3428 // Call allocation profiler 3429 AllocationProfiler::iterate_since_last_gc(); 3430 // Fill TLAB's and such 3431 ensure_parsability(true); 3432 } 3433 3434 void G1CollectedHeap::gc_epilogue(bool full /* Ignored */) { 3435 // FIXME: what is this about? 3436 // I'm ignoring the "fill_newgen()" call if "alloc_event_enabled" 3437 // is set. 3438 COMPILER2_PRESENT(assert(DerivedPointerTable::is_empty(), 3439 "derived pointer present")); 3440 // always_do_update_barrier = true; 3441 3442 // We have just completed a GC. Update the soft reference 3443 // policy with the new heap occupancy 3444 Universe::update_heap_info_at_gc(); 3445 } 3446 3447 HeapWord* G1CollectedHeap::do_collection_pause(size_t word_size, 3448 unsigned int gc_count_before, 3449 bool* succeeded) { 3450 assert_heap_not_locked_and_not_at_safepoint(); 3451 g1_policy()->record_stop_world_start(); 3452 VM_G1IncCollectionPause op(gc_count_before, 3453 word_size, 3454 false, /* should_initiate_conc_mark */ 3455 g1_policy()->max_pause_time_ms(), 3456 GCCause::_g1_inc_collection_pause); 3457 VMThread::execute(&op); 3458 3459 HeapWord* result = op.result(); 3460 bool ret_succeeded = op.prologue_succeeded() && op.pause_succeeded(); 3461 assert(result == NULL || ret_succeeded, 3462 "the result should be NULL if the VM did not succeed"); 3463 *succeeded = ret_succeeded; 3464 3465 assert_heap_not_locked(); 3466 return result; 3467 } 3468 3469 void 3470 G1CollectedHeap::doConcurrentMark() { 3471 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); 3472 if (!_cmThread->in_progress()) { 3473 _cmThread->set_started(); 3474 CGC_lock->notify(); 3475 } 3476 } 3477 3478 size_t G1CollectedHeap::pending_card_num() { 3479 size_t extra_cards = 0; 3480 JavaThread *curr = Threads::first(); 3481 while (curr != NULL) { 3482 DirtyCardQueue& dcq = curr->dirty_card_queue(); 3483 extra_cards += dcq.size(); 3484 curr = curr->next(); 3485 } 3486 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3487 size_t buffer_size = dcqs.buffer_size(); 3488 size_t buffer_num = dcqs.completed_buffers_num(); 3489 return buffer_size * buffer_num + extra_cards; 3490 } 3491 3492 size_t G1CollectedHeap::max_pending_card_num() { 3493 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set(); 3494 size_t buffer_size = dcqs.buffer_size(); 3495 size_t buffer_num = dcqs.completed_buffers_num(); 3496 int thread_num = Threads::number_of_threads(); 3497 return (buffer_num + thread_num) * buffer_size; 3498 } 3499 3500 size_t G1CollectedHeap::cards_scanned() { 3501 return g1_rem_set()->cardsScanned(); 3502 } 3503 3504 void 3505 G1CollectedHeap::setup_surviving_young_words() { 3506 assert(_surviving_young_words == NULL, "pre-condition"); 3507 uint array_length = g1_policy()->young_cset_region_length(); 3508 _surviving_young_words = NEW_C_HEAP_ARRAY(size_t, (size_t) array_length, mtGC); 3509 if (_surviving_young_words == NULL) { 3510 vm_exit_out_of_memory(sizeof(size_t) * array_length, 3511 "Not enough space for young surv words summary."); 3512 } 3513 memset(_surviving_young_words, 0, (size_t) array_length * sizeof(size_t)); 3514 #ifdef ASSERT 3515 for (uint i = 0; i < array_length; ++i) { 3516 assert( _surviving_young_words[i] == 0, "memset above" ); 3517 } 3518 #endif // !ASSERT 3519 } 3520 3521 void 3522 G1CollectedHeap::update_surviving_young_words(size_t* surv_young_words) { 3523 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); 3524 uint array_length = g1_policy()->young_cset_region_length(); 3525 for (uint i = 0; i < array_length; ++i) { 3526 _surviving_young_words[i] += surv_young_words[i]; 3527 } 3528 } 3529 3530 void 3531 G1CollectedHeap::cleanup_surviving_young_words() { 3532 guarantee( _surviving_young_words != NULL, "pre-condition" ); 3533 FREE_C_HEAP_ARRAY(size_t, _surviving_young_words, mtGC); 3534 _surviving_young_words = NULL; 3535 } 3536 3537 #ifdef ASSERT 3538 class VerifyCSetClosure: public HeapRegionClosure { 3539 public: 3540 bool doHeapRegion(HeapRegion* hr) { 3541 // Here we check that the CSet region's RSet is ready for parallel 3542 // iteration. The fields that we'll verify are only manipulated 3543 // when the region is part of a CSet and is collected. Afterwards, 3544 // we reset these fields when we clear the region's RSet (when the 3545 // region is freed) so they are ready when the region is 3546 // re-allocated. The only exception to this is if there's an 3547 // evacuation failure and instead of freeing the region we leave 3548 // it in the heap. In that case, we reset these fields during 3549 // evacuation failure handling. 3550 guarantee(hr->rem_set()->verify_ready_for_par_iteration(), "verification"); 3551 3552 // Here's a good place to add any other checks we'd like to 3553 // perform on CSet regions. 3554 return false; 3555 } 3556 }; 3557 #endif // ASSERT 3558 3559 #if TASKQUEUE_STATS 3560 void G1CollectedHeap::print_taskqueue_stats_hdr(outputStream* const st) { 3561 st->print_raw_cr("GC Task Stats"); 3562 st->print_raw("thr "); TaskQueueStats::print_header(1, st); st->cr(); 3563 st->print_raw("--- "); TaskQueueStats::print_header(2, st); st->cr(); 3564 } 3565 3566 void G1CollectedHeap::print_taskqueue_stats(outputStream* const st) const { 3567 print_taskqueue_stats_hdr(st); 3568 3569 TaskQueueStats totals; 3570 const int n = workers() != NULL ? workers()->total_workers() : 1; 3571 for (int i = 0; i < n; ++i) { 3572 st->print("%3d ", i); task_queue(i)->stats.print(st); st->cr(); 3573 totals += task_queue(i)->stats; 3574 } 3575 st->print_raw("tot "); totals.print(st); st->cr(); 3576 3577 DEBUG_ONLY(totals.verify()); 3578 } 3579 3580 void G1CollectedHeap::reset_taskqueue_stats() { 3581 const int n = workers() != NULL ? workers()->total_workers() : 1; 3582 for (int i = 0; i < n; ++i) { 3583 task_queue(i)->stats.reset(); 3584 } 3585 } 3586 #endif // TASKQUEUE_STATS 3587 3588 bool 3589 G1CollectedHeap::do_collection_pause_at_safepoint(double target_pause_time_ms) { 3590 assert_at_safepoint(true /* should_be_vm_thread */); 3591 guarantee(!is_gc_active(), "collection is not reentrant"); 3592 3593 if (GC_locker::check_active_before_gc()) { 3594 return false; 3595 } 3596 3597 SvcGCMarker sgcm(SvcGCMarker::MINOR); 3598 ResourceMark rm; 3599 3600 print_heap_before_gc(); 3601 3602 HRSPhaseSetter x(HRSPhaseEvacuation); 3603 verify_region_sets_optional(); 3604 verify_dirty_young_regions(); 3605 3606 // This call will decide whether this pause is an initial-mark 3607 // pause. If it is, during_initial_mark_pause() will return true 3608 // for the duration of this pause. 3609 g1_policy()->decide_on_conc_mark_initiation(); 3610 3611 // We do not allow initial-mark to be piggy-backed on a mixed GC. 3612 assert(!g1_policy()->during_initial_mark_pause() || 3613 g1_policy()->gcs_are_young(), "sanity"); 3614 3615 // We also do not allow mixed GCs during marking. 3616 assert(!mark_in_progress() || g1_policy()->gcs_are_young(), "sanity"); 3617 3618 // Record whether this pause is an initial mark. When the current 3619 // thread has completed its logging output and it's safe to signal 3620 // the CM thread, the flag's value in the policy has been reset. 3621 bool should_start_conc_mark = g1_policy()->during_initial_mark_pause(); 3622 3623 // Inner scope for scope based logging, timers, and stats collection 3624 { 3625 if (g1_policy()->during_initial_mark_pause()) { 3626 // We are about to start a marking cycle, so we increment the 3627 // full collection counter. 3628 increment_old_marking_cycles_started(); 3629 } 3630 // if the log level is "finer" is on, we'll print long statistics information 3631 // in the collector policy code, so let's not print this as the output 3632 // is messy if we do. 3633 gclog_or_tty->date_stamp(G1Log::fine() && PrintGCDateStamps); 3634 TraceCPUTime tcpu(G1Log::finer(), true, gclog_or_tty); 3635 3636 int active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 3637 workers()->active_workers() : 1); 3638 g1_policy()->phase_times()->note_gc_start(os::elapsedTime(), active_workers, 3639 g1_policy()->gcs_are_young(), g1_policy()->during_initial_mark_pause(), gc_cause()); 3640 3641 TraceCollectorStats tcs(g1mm()->incremental_collection_counters()); 3642 TraceMemoryManagerStats tms(false /* fullGC */, gc_cause()); 3643 3644 // If the secondary_free_list is not empty, append it to the 3645 // free_list. No need to wait for the cleanup operation to finish; 3646 // the region allocation code will check the secondary_free_list 3647 // and wait if necessary. If the G1StressConcRegionFreeing flag is 3648 // set, skip this step so that the region allocation code has to 3649 // get entries from the secondary_free_list. 3650 if (!G1StressConcRegionFreeing) { 3651 append_secondary_free_list_if_not_empty_with_lock(); 3652 } 3653 3654 assert(check_young_list_well_formed(), 3655 "young list should be well formed"); 3656 3657 // Don't dynamically change the number of GC threads this early. A value of 3658 // 0 is used to indicate serial work. When parallel work is done, 3659 // it will be set. 3660 3661 { // Call to jvmpi::post_class_unload_events must occur outside of active GC 3662 IsGCActiveMark x; 3663 3664 gc_prologue(false); 3665 increment_total_collections(false /* full gc */); 3666 increment_gc_time_stamp(); 3667 3668 if (VerifyBeforeGC && total_collections() >= VerifyGCStartAt) { 3669 HandleMark hm; // Discard invalid handles created during verification 3670 gclog_or_tty->print(" VerifyBeforeGC:"); 3671 prepare_for_verify(); 3672 Universe::verify(/* silent */ false, 3673 /* option */ VerifyOption_G1UsePrevMarking); 3674 } 3675 3676 COMPILER2_PRESENT(DerivedPointerTable::clear()); 3677 3678 // Please see comment in g1CollectedHeap.hpp and 3679 // G1CollectedHeap::ref_processing_init() to see how 3680 // reference processing currently works in G1. 3681 3682 // Enable discovery in the STW reference processor 3683 ref_processor_stw()->enable_discovery(true /*verify_disabled*/, 3684 true /*verify_no_refs*/); 3685 3686 { 3687 // We want to temporarily turn off discovery by the 3688 // CM ref processor, if necessary, and turn it back on 3689 // on again later if we do. Using a scoped 3690 // NoRefDiscovery object will do this. 3691 NoRefDiscovery no_cm_discovery(ref_processor_cm()); 3692 3693 // Forget the current alloc region (we might even choose it to be part 3694 // of the collection set!). 3695 release_mutator_alloc_region(); 3696 3697 // We should call this after we retire the mutator alloc 3698 // region(s) so that all the ALLOC / RETIRE events are generated 3699 // before the start GC event. 3700 _hr_printer.start_gc(false /* full */, (size_t) total_collections()); 3701 3702 // This timing is only used by the ergonomics to handle our pause target. 3703 // It is unclear why this should not include the full pause. We will 3704 // investigate this in CR 7178365. 3705 // 3706 // Preserving the old comment here if that helps the investigation: 3707 // 3708 // The elapsed time induced by the start time below deliberately elides 3709 // the possible verification above. 3710 double sample_start_time_sec = os::elapsedTime(); 3711 size_t start_used_bytes = used(); 3712 3713 #if YOUNG_LIST_VERBOSE 3714 gclog_or_tty->print_cr("\nBefore recording pause start.\nYoung_list:"); 3715 _young_list->print(); 3716 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3717 #endif // YOUNG_LIST_VERBOSE 3718 3719 g1_policy()->record_collection_pause_start(sample_start_time_sec, 3720 start_used_bytes); 3721 3722 double scan_wait_start = os::elapsedTime(); 3723 // We have to wait until the CM threads finish scanning the 3724 // root regions as it's the only way to ensure that all the 3725 // objects on them have been correctly scanned before we start 3726 // moving them during the GC. 3727 bool waited = _cm->root_regions()->wait_until_scan_finished(); 3728 double wait_time_ms = 0.0; 3729 if (waited) { 3730 double scan_wait_end = os::elapsedTime(); 3731 wait_time_ms = (scan_wait_end - scan_wait_start) * 1000.0; 3732 } 3733 g1_policy()->phase_times()->record_root_region_scan_wait_time(wait_time_ms); 3734 3735 #if YOUNG_LIST_VERBOSE 3736 gclog_or_tty->print_cr("\nAfter recording pause start.\nYoung_list:"); 3737 _young_list->print(); 3738 #endif // YOUNG_LIST_VERBOSE 3739 3740 if (g1_policy()->during_initial_mark_pause()) { 3741 concurrent_mark()->checkpointRootsInitialPre(); 3742 } 3743 perm_gen()->save_marks(); 3744 3745 #if YOUNG_LIST_VERBOSE 3746 gclog_or_tty->print_cr("\nBefore choosing collection set.\nYoung_list:"); 3747 _young_list->print(); 3748 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3749 #endif // YOUNG_LIST_VERBOSE 3750 3751 g1_policy()->finalize_cset(target_pause_time_ms); 3752 3753 _cm->note_start_of_gc(); 3754 // We should not verify the per-thread SATB buffers given that 3755 // we have not filtered them yet (we'll do so during the 3756 // GC). We also call this after finalize_cset() to 3757 // ensure that the CSet has been finalized. 3758 _cm->verify_no_cset_oops(true /* verify_stacks */, 3759 true /* verify_enqueued_buffers */, 3760 false /* verify_thread_buffers */, 3761 true /* verify_fingers */); 3762 3763 if (_hr_printer.is_active()) { 3764 HeapRegion* hr = g1_policy()->collection_set(); 3765 while (hr != NULL) { 3766 G1HRPrinter::RegionType type; 3767 if (!hr->is_young()) { 3768 type = G1HRPrinter::Old; 3769 } else if (hr->is_survivor()) { 3770 type = G1HRPrinter::Survivor; 3771 } else { 3772 type = G1HRPrinter::Eden; 3773 } 3774 _hr_printer.cset(hr); 3775 hr = hr->next_in_collection_set(); 3776 } 3777 } 3778 3779 #ifdef ASSERT 3780 VerifyCSetClosure cl; 3781 collection_set_iterate(&cl); 3782 #endif // ASSERT 3783 3784 setup_surviving_young_words(); 3785 3786 // Initialize the GC alloc regions. 3787 init_gc_alloc_regions(); 3788 3789 // Actually do the work... 3790 evacuate_collection_set(); 3791 3792 // We do this to mainly verify the per-thread SATB buffers 3793 // (which have been filtered by now) since we didn't verify 3794 // them earlier. No point in re-checking the stacks / enqueued 3795 // buffers given that the CSet has not changed since last time 3796 // we checked. 3797 _cm->verify_no_cset_oops(false /* verify_stacks */, 3798 false /* verify_enqueued_buffers */, 3799 true /* verify_thread_buffers */, 3800 true /* verify_fingers */); 3801 3802 free_collection_set(g1_policy()->collection_set()); 3803 g1_policy()->clear_collection_set(); 3804 3805 cleanup_surviving_young_words(); 3806 3807 // Start a new incremental collection set for the next pause. 3808 g1_policy()->start_incremental_cset_building(); 3809 3810 // Clear the _cset_fast_test bitmap in anticipation of adding 3811 // regions to the incremental collection set for the next 3812 // evacuation pause. 3813 clear_cset_fast_test(); 3814 3815 _young_list->reset_sampled_info(); 3816 3817 // Don't check the whole heap at this point as the 3818 // GC alloc regions from this pause have been tagged 3819 // as survivors and moved on to the survivor list. 3820 // Survivor regions will fail the !is_young() check. 3821 assert(check_young_list_empty(false /* check_heap */), 3822 "young list should be empty"); 3823 3824 #if YOUNG_LIST_VERBOSE 3825 gclog_or_tty->print_cr("Before recording survivors.\nYoung List:"); 3826 _young_list->print(); 3827 #endif // YOUNG_LIST_VERBOSE 3828 3829 g1_policy()->record_survivor_regions(_young_list->survivor_length(), 3830 _young_list->first_survivor_region(), 3831 _young_list->last_survivor_region()); 3832 3833 _young_list->reset_auxilary_lists(); 3834 3835 if (evacuation_failed()) { 3836 _summary_bytes_used = recalculate_used(); 3837 } else { 3838 // The "used" of the the collection set have already been subtracted 3839 // when they were freed. Add in the bytes evacuated. 3840 _summary_bytes_used += g1_policy()->bytes_copied_during_gc(); 3841 } 3842 3843 if (g1_policy()->during_initial_mark_pause()) { 3844 // We have to do this before we notify the CM threads that 3845 // they can start working to make sure that all the 3846 // appropriate initialization is done on the CM object. 3847 concurrent_mark()->checkpointRootsInitialPost(); 3848 set_marking_started(); 3849 // Note that we don't actually trigger the CM thread at 3850 // this point. We do that later when we're sure that 3851 // the current thread has completed its logging output. 3852 } 3853 3854 allocate_dummy_regions(); 3855 3856 #if YOUNG_LIST_VERBOSE 3857 gclog_or_tty->print_cr("\nEnd of the pause.\nYoung_list:"); 3858 _young_list->print(); 3859 g1_policy()->print_collection_set(g1_policy()->inc_cset_head(), gclog_or_tty); 3860 #endif // YOUNG_LIST_VERBOSE 3861 3862 init_mutator_alloc_region(); 3863 3864 { 3865 size_t expand_bytes = g1_policy()->expansion_amount(); 3866 if (expand_bytes > 0) { 3867 size_t bytes_before = capacity(); 3868 // No need for an ergo verbose message here, 3869 // expansion_amount() does this when it returns a value > 0. 3870 if (!expand(expand_bytes)) { 3871 // We failed to expand the heap so let's verify that 3872 // committed/uncommitted amount match the backing store 3873 assert(capacity() == _g1_storage.committed_size(), "committed size mismatch"); 3874 assert(max_capacity() == _g1_storage.reserved_size(), "reserved size mismatch"); 3875 } 3876 } 3877 } 3878 3879 // We redo the verificaiton but now wrt to the new CSet which 3880 // has just got initialized after the previous CSet was freed. 3881 _cm->verify_no_cset_oops(true /* verify_stacks */, 3882 true /* verify_enqueued_buffers */, 3883 true /* verify_thread_buffers */, 3884 true /* verify_fingers */); 3885 _cm->note_end_of_gc(); 3886 3887 // This timing is only used by the ergonomics to handle our pause target. 3888 // It is unclear why this should not include the full pause. We will 3889 // investigate this in CR 7178365. 3890 double sample_end_time_sec = os::elapsedTime(); 3891 double pause_time_ms = (sample_end_time_sec - sample_start_time_sec) * MILLIUNITS; 3892 g1_policy()->record_collection_pause_end(pause_time_ms); 3893 3894 MemoryService::track_memory_usage(); 3895 3896 // In prepare_for_verify() below we'll need to scan the deferred 3897 // update buffers to bring the RSets up-to-date if 3898 // G1HRRSFlushLogBuffersOnVerify has been set. While scanning 3899 // the update buffers we'll probably need to scan cards on the 3900 // regions we just allocated to (i.e., the GC alloc 3901 // regions). However, during the last GC we called 3902 // set_saved_mark() on all the GC alloc regions, so card 3903 // scanning might skip the [saved_mark_word()...top()] area of 3904 // those regions (i.e., the area we allocated objects into 3905 // during the last GC). But it shouldn't. Given that 3906 // saved_mark_word() is conditional on whether the GC time stamp 3907 // on the region is current or not, by incrementing the GC time 3908 // stamp here we invalidate all the GC time stamps on all the 3909 // regions and saved_mark_word() will simply return top() for 3910 // all the regions. This is a nicer way of ensuring this rather 3911 // than iterating over the regions and fixing them. In fact, the 3912 // GC time stamp increment here also ensures that 3913 // saved_mark_word() will return top() between pauses, i.e., 3914 // during concurrent refinement. So we don't need the 3915 // is_gc_active() check to decided which top to use when 3916 // scanning cards (see CR 7039627). 3917 increment_gc_time_stamp(); 3918 3919 if (VerifyAfterGC && total_collections() >= VerifyGCStartAt) { 3920 HandleMark hm; // Discard invalid handles created during verification 3921 gclog_or_tty->print(" VerifyAfterGC:"); 3922 prepare_for_verify(); 3923 Universe::verify(/* silent */ false, 3924 /* option */ VerifyOption_G1UsePrevMarking); 3925 } 3926 3927 assert(!ref_processor_stw()->discovery_enabled(), "Postcondition"); 3928 ref_processor_stw()->verify_no_references_recorded(); 3929 3930 // CM reference discovery will be re-enabled if necessary. 3931 } 3932 3933 // We should do this after we potentially expand the heap so 3934 // that all the COMMIT events are generated before the end GC 3935 // event, and after we retire the GC alloc regions so that all 3936 // RETIRE events are generated before the end GC event. 3937 _hr_printer.end_gc(false /* full */, (size_t) total_collections()); 3938 3939 if (mark_in_progress()) { 3940 concurrent_mark()->update_g1_committed(); 3941 } 3942 3943 #ifdef TRACESPINNING 3944 ParallelTaskTerminator::print_termination_counts(); 3945 #endif 3946 3947 gc_epilogue(false); 3948 3949 g1_policy()->phase_times()->note_gc_end(os::elapsedTime()); 3950 3951 // We have to do this after we decide whether to expand the heap or not. 3952 g1_policy()->print_heap_transition(); 3953 } 3954 3955 // It is not yet to safe to tell the concurrent mark to 3956 // start as we have some optional output below. We don't want the 3957 // output from the concurrent mark thread interfering with this 3958 // logging output either. 3959 3960 _hrs.verify_optional(); 3961 verify_region_sets_optional(); 3962 3963 TASKQUEUE_STATS_ONLY(if (ParallelGCVerbose) print_taskqueue_stats()); 3964 TASKQUEUE_STATS_ONLY(reset_taskqueue_stats()); 3965 3966 print_heap_after_gc(); 3967 3968 // We must call G1MonitoringSupport::update_sizes() in the same scoping level 3969 // as an active TraceMemoryManagerStats object (i.e. before the destructor for the 3970 // TraceMemoryManagerStats is called) so that the G1 memory pools are updated 3971 // before any GC notifications are raised. 3972 g1mm()->update_sizes(); 3973 } 3974 3975 if (G1SummarizeRSetStats && 3976 (G1SummarizeRSetStatsPeriod > 0) && 3977 (total_collections() % G1SummarizeRSetStatsPeriod == 0)) { 3978 g1_rem_set()->print_summary_info(); 3979 } 3980 3981 // It should now be safe to tell the concurrent mark thread to start 3982 // without its logging output interfering with the logging output 3983 // that came from the pause. 3984 3985 if (should_start_conc_mark) { 3986 // CAUTION: after the doConcurrentMark() call below, 3987 // the concurrent marking thread(s) could be running 3988 // concurrently with us. Make sure that anything after 3989 // this point does not assume that we are the only GC thread 3990 // running. Note: of course, the actual marking work will 3991 // not start until the safepoint itself is released in 3992 // ConcurrentGCThread::safepoint_desynchronize(). 3993 doConcurrentMark(); 3994 } 3995 3996 return true; 3997 } 3998 3999 size_t G1CollectedHeap::desired_plab_sz(GCAllocPurpose purpose) 4000 { 4001 size_t gclab_word_size; 4002 switch (purpose) { 4003 case GCAllocForSurvived: 4004 gclab_word_size = YoungPLABSize; 4005 break; 4006 case GCAllocForTenured: 4007 gclab_word_size = OldPLABSize; 4008 break; 4009 default: 4010 assert(false, "unknown GCAllocPurpose"); 4011 gclab_word_size = OldPLABSize; 4012 break; 4013 } 4014 return gclab_word_size; 4015 } 4016 4017 void G1CollectedHeap::init_mutator_alloc_region() { 4018 assert(_mutator_alloc_region.get() == NULL, "pre-condition"); 4019 _mutator_alloc_region.init(); 4020 } 4021 4022 void G1CollectedHeap::release_mutator_alloc_region() { 4023 _mutator_alloc_region.release(); 4024 assert(_mutator_alloc_region.get() == NULL, "post-condition"); 4025 } 4026 4027 void G1CollectedHeap::init_gc_alloc_regions() { 4028 assert_at_safepoint(true /* should_be_vm_thread */); 4029 4030 _survivor_gc_alloc_region.init(); 4031 _old_gc_alloc_region.init(); 4032 HeapRegion* retained_region = _retained_old_gc_alloc_region; 4033 _retained_old_gc_alloc_region = NULL; 4034 4035 // We will discard the current GC alloc region if: 4036 // a) it's in the collection set (it can happen!), 4037 // b) it's already full (no point in using it), 4038 // c) it's empty (this means that it was emptied during 4039 // a cleanup and it should be on the free list now), or 4040 // d) it's humongous (this means that it was emptied 4041 // during a cleanup and was added to the free list, but 4042 // has been subseqently used to allocate a humongous 4043 // object that may be less than the region size). 4044 if (retained_region != NULL && 4045 !retained_region->in_collection_set() && 4046 !(retained_region->top() == retained_region->end()) && 4047 !retained_region->is_empty() && 4048 !retained_region->isHumongous()) { 4049 retained_region->set_saved_mark(); 4050 // The retained region was added to the old region set when it was 4051 // retired. We have to remove it now, since we don't allow regions 4052 // we allocate to in the region sets. We'll re-add it later, when 4053 // it's retired again. 4054 _old_set.remove(retained_region); 4055 bool during_im = g1_policy()->during_initial_mark_pause(); 4056 retained_region->note_start_of_copying(during_im); 4057 _old_gc_alloc_region.set(retained_region); 4058 _hr_printer.reuse(retained_region); 4059 } 4060 } 4061 4062 void G1CollectedHeap::release_gc_alloc_regions() { 4063 _survivor_gc_alloc_region.release(); 4064 // If we have an old GC alloc region to release, we'll save it in 4065 // _retained_old_gc_alloc_region. If we don't 4066 // _retained_old_gc_alloc_region will become NULL. This is what we 4067 // want either way so no reason to check explicitly for either 4068 // condition. 4069 _retained_old_gc_alloc_region = _old_gc_alloc_region.release(); 4070 } 4071 4072 void G1CollectedHeap::abandon_gc_alloc_regions() { 4073 assert(_survivor_gc_alloc_region.get() == NULL, "pre-condition"); 4074 assert(_old_gc_alloc_region.get() == NULL, "pre-condition"); 4075 _retained_old_gc_alloc_region = NULL; 4076 } 4077 4078 void G1CollectedHeap::init_for_evac_failure(OopsInHeapRegionClosure* cl) { 4079 _drain_in_progress = false; 4080 set_evac_failure_closure(cl); 4081 _evac_failure_scan_stack = new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true); 4082 } 4083 4084 void G1CollectedHeap::finalize_for_evac_failure() { 4085 assert(_evac_failure_scan_stack != NULL && 4086 _evac_failure_scan_stack->length() == 0, 4087 "Postcondition"); 4088 assert(!_drain_in_progress, "Postcondition"); 4089 delete _evac_failure_scan_stack; 4090 _evac_failure_scan_stack = NULL; 4091 } 4092 4093 void G1CollectedHeap::remove_self_forwarding_pointers() { 4094 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4095 4096 G1ParRemoveSelfForwardPtrsTask rsfp_task(this); 4097 4098 if (G1CollectedHeap::use_parallel_gc_threads()) { 4099 set_par_threads(); 4100 workers()->run_task(&rsfp_task); 4101 set_par_threads(0); 4102 } else { 4103 rsfp_task.work(0); 4104 } 4105 4106 assert(check_cset_heap_region_claim_values(HeapRegion::ParEvacFailureClaimValue), "sanity"); 4107 4108 // Reset the claim values in the regions in the collection set. 4109 reset_cset_heap_region_claim_values(); 4110 4111 assert(check_cset_heap_region_claim_values(HeapRegion::InitialClaimValue), "sanity"); 4112 4113 // Now restore saved marks, if any. 4114 if (_objs_with_preserved_marks != NULL) { 4115 assert(_preserved_marks_of_objs != NULL, "Both or none."); 4116 guarantee(_objs_with_preserved_marks->length() == 4117 _preserved_marks_of_objs->length(), "Both or none."); 4118 for (int i = 0; i < _objs_with_preserved_marks->length(); i++) { 4119 oop obj = _objs_with_preserved_marks->at(i); 4120 markOop m = _preserved_marks_of_objs->at(i); 4121 obj->set_mark(m); 4122 } 4123 4124 // Delete the preserved marks growable arrays (allocated on the C heap). 4125 delete _objs_with_preserved_marks; 4126 delete _preserved_marks_of_objs; 4127 _objs_with_preserved_marks = NULL; 4128 _preserved_marks_of_objs = NULL; 4129 } 4130 } 4131 4132 void G1CollectedHeap::push_on_evac_failure_scan_stack(oop obj) { 4133 _evac_failure_scan_stack->push(obj); 4134 } 4135 4136 void G1CollectedHeap::drain_evac_failure_scan_stack() { 4137 assert(_evac_failure_scan_stack != NULL, "precondition"); 4138 4139 while (_evac_failure_scan_stack->length() > 0) { 4140 oop obj = _evac_failure_scan_stack->pop(); 4141 _evac_failure_closure->set_region(heap_region_containing(obj)); 4142 obj->oop_iterate_backwards(_evac_failure_closure); 4143 } 4144 } 4145 4146 oop 4147 G1CollectedHeap::handle_evacuation_failure_par(OopsInHeapRegionClosure* cl, 4148 oop old) { 4149 assert(obj_in_cs(old), 4150 err_msg("obj: "PTR_FORMAT" should still be in the CSet", 4151 (HeapWord*) old)); 4152 markOop m = old->mark(); 4153 oop forward_ptr = old->forward_to_atomic(old); 4154 if (forward_ptr == NULL) { 4155 // Forward-to-self succeeded. 4156 4157 if (_evac_failure_closure != cl) { 4158 MutexLockerEx x(EvacFailureStack_lock, Mutex::_no_safepoint_check_flag); 4159 assert(!_drain_in_progress, 4160 "Should only be true while someone holds the lock."); 4161 // Set the global evac-failure closure to the current thread's. 4162 assert(_evac_failure_closure == NULL, "Or locking has failed."); 4163 set_evac_failure_closure(cl); 4164 // Now do the common part. 4165 handle_evacuation_failure_common(old, m); 4166 // Reset to NULL. 4167 set_evac_failure_closure(NULL); 4168 } else { 4169 // The lock is already held, and this is recursive. 4170 assert(_drain_in_progress, "This should only be the recursive case."); 4171 handle_evacuation_failure_common(old, m); 4172 } 4173 return old; 4174 } else { 4175 // Forward-to-self failed. Either someone else managed to allocate 4176 // space for this object (old != forward_ptr) or they beat us in 4177 // self-forwarding it (old == forward_ptr). 4178 assert(old == forward_ptr || !obj_in_cs(forward_ptr), 4179 err_msg("obj: "PTR_FORMAT" forwarded to: "PTR_FORMAT" " 4180 "should not be in the CSet", 4181 (HeapWord*) old, (HeapWord*) forward_ptr)); 4182 return forward_ptr; 4183 } 4184 } 4185 4186 void G1CollectedHeap::handle_evacuation_failure_common(oop old, markOop m) { 4187 set_evacuation_failed(true); 4188 4189 preserve_mark_if_necessary(old, m); 4190 4191 HeapRegion* r = heap_region_containing(old); 4192 if (!r->evacuation_failed()) { 4193 r->set_evacuation_failed(true); 4194 _hr_printer.evac_failure(r); 4195 } 4196 4197 push_on_evac_failure_scan_stack(old); 4198 4199 if (!_drain_in_progress) { 4200 // prevent recursion in copy_to_survivor_space() 4201 _drain_in_progress = true; 4202 drain_evac_failure_scan_stack(); 4203 _drain_in_progress = false; 4204 } 4205 } 4206 4207 void G1CollectedHeap::preserve_mark_if_necessary(oop obj, markOop m) { 4208 assert(evacuation_failed(), "Oversaving!"); 4209 // We want to call the "for_promotion_failure" version only in the 4210 // case of a promotion failure. 4211 if (m->must_be_preserved_for_promotion_failure(obj)) { 4212 if (_objs_with_preserved_marks == NULL) { 4213 assert(_preserved_marks_of_objs == NULL, "Both or none."); 4214 _objs_with_preserved_marks = 4215 new (ResourceObj::C_HEAP, mtGC) GrowableArray<oop>(40, true); 4216 _preserved_marks_of_objs = 4217 new (ResourceObj::C_HEAP, mtGC) GrowableArray<markOop>(40, true); 4218 } 4219 _objs_with_preserved_marks->push(obj); 4220 _preserved_marks_of_objs->push(m); 4221 } 4222 } 4223 4224 HeapWord* G1CollectedHeap::par_allocate_during_gc(GCAllocPurpose purpose, 4225 size_t word_size) { 4226 if (purpose == GCAllocForSurvived) { 4227 HeapWord* result = survivor_attempt_allocation(word_size); 4228 if (result != NULL) { 4229 return result; 4230 } else { 4231 // Let's try to allocate in the old gen in case we can fit the 4232 // object there. 4233 return old_attempt_allocation(word_size); 4234 } 4235 } else { 4236 assert(purpose == GCAllocForTenured, "sanity"); 4237 HeapWord* result = old_attempt_allocation(word_size); 4238 if (result != NULL) { 4239 return result; 4240 } else { 4241 // Let's try to allocate in the survivors in case we can fit the 4242 // object there. 4243 return survivor_attempt_allocation(word_size); 4244 } 4245 } 4246 4247 ShouldNotReachHere(); 4248 // Trying to keep some compilers happy. 4249 return NULL; 4250 } 4251 4252 G1ParGCAllocBuffer::G1ParGCAllocBuffer(size_t gclab_word_size) : 4253 ParGCAllocBuffer(gclab_word_size), _retired(false) { } 4254 4255 G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num) 4256 : _g1h(g1h), 4257 _refs(g1h->task_queue(queue_num)), 4258 _dcq(&g1h->dirty_card_queue_set()), 4259 _ct_bs((CardTableModRefBS*)_g1h->barrier_set()), 4260 _g1_rem(g1h->g1_rem_set()), 4261 _hash_seed(17), _queue_num(queue_num), 4262 _term_attempts(0), 4263 _surviving_alloc_buffer(g1h->desired_plab_sz(GCAllocForSurvived)), 4264 _tenured_alloc_buffer(g1h->desired_plab_sz(GCAllocForTenured)), 4265 _age_table(false), 4266 _strong_roots_time(0), _term_time(0), 4267 _alloc_buffer_waste(0), _undo_waste(0) { 4268 // we allocate G1YoungSurvRateNumRegions plus one entries, since 4269 // we "sacrifice" entry 0 to keep track of surviving bytes for 4270 // non-young regions (where the age is -1) 4271 // We also add a few elements at the beginning and at the end in 4272 // an attempt to eliminate cache contention 4273 uint real_length = 1 + _g1h->g1_policy()->young_cset_region_length(); 4274 uint array_length = PADDING_ELEM_NUM + 4275 real_length + 4276 PADDING_ELEM_NUM; 4277 _surviving_young_words_base = NEW_C_HEAP_ARRAY(size_t, array_length, mtGC); 4278 if (_surviving_young_words_base == NULL) 4279 vm_exit_out_of_memory(array_length * sizeof(size_t), 4280 "Not enough space for young surv histo."); 4281 _surviving_young_words = _surviving_young_words_base + PADDING_ELEM_NUM; 4282 memset(_surviving_young_words, 0, (size_t) real_length * sizeof(size_t)); 4283 4284 _alloc_buffers[GCAllocForSurvived] = &_surviving_alloc_buffer; 4285 _alloc_buffers[GCAllocForTenured] = &_tenured_alloc_buffer; 4286 4287 _start = os::elapsedTime(); 4288 } 4289 4290 void 4291 G1ParScanThreadState::print_termination_stats_hdr(outputStream* const st) 4292 { 4293 st->print_raw_cr("GC Termination Stats"); 4294 st->print_raw_cr(" elapsed --strong roots-- -------termination-------" 4295 " ------waste (KiB)------"); 4296 st->print_raw_cr("thr ms ms % ms % attempts" 4297 " total alloc undo"); 4298 st->print_raw_cr("--- --------- --------- ------ --------- ------ --------" 4299 " ------- ------- -------"); 4300 } 4301 4302 void 4303 G1ParScanThreadState::print_termination_stats(int i, 4304 outputStream* const st) const 4305 { 4306 const double elapsed_ms = elapsed_time() * 1000.0; 4307 const double s_roots_ms = strong_roots_time() * 1000.0; 4308 const double term_ms = term_time() * 1000.0; 4309 st->print_cr("%3d %9.2f %9.2f %6.2f " 4310 "%9.2f %6.2f " SIZE_FORMAT_W(8) " " 4311 SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7) " " SIZE_FORMAT_W(7), 4312 i, elapsed_ms, s_roots_ms, s_roots_ms * 100 / elapsed_ms, 4313 term_ms, term_ms * 100 / elapsed_ms, term_attempts(), 4314 (alloc_buffer_waste() + undo_waste()) * HeapWordSize / K, 4315 alloc_buffer_waste() * HeapWordSize / K, 4316 undo_waste() * HeapWordSize / K); 4317 } 4318 4319 #ifdef ASSERT 4320 bool G1ParScanThreadState::verify_ref(narrowOop* ref) const { 4321 assert(ref != NULL, "invariant"); 4322 assert(UseCompressedOops, "sanity"); 4323 assert(!has_partial_array_mask(ref), err_msg("ref=" PTR_FORMAT, ref)); 4324 oop p = oopDesc::load_decode_heap_oop(ref); 4325 assert(_g1h->is_in_g1_reserved(p), 4326 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4327 return true; 4328 } 4329 4330 bool G1ParScanThreadState::verify_ref(oop* ref) const { 4331 assert(ref != NULL, "invariant"); 4332 if (has_partial_array_mask(ref)) { 4333 // Must be in the collection set--it's already been copied. 4334 oop p = clear_partial_array_mask(ref); 4335 assert(_g1h->obj_in_cs(p), 4336 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4337 } else { 4338 oop p = oopDesc::load_decode_heap_oop(ref); 4339 assert(_g1h->is_in_g1_reserved(p), 4340 err_msg("ref=" PTR_FORMAT " p=" PTR_FORMAT, ref, intptr_t(p))); 4341 } 4342 return true; 4343 } 4344 4345 bool G1ParScanThreadState::verify_task(StarTask ref) const { 4346 if (ref.is_narrow()) { 4347 return verify_ref((narrowOop*) ref); 4348 } else { 4349 return verify_ref((oop*) ref); 4350 } 4351 } 4352 #endif // ASSERT 4353 4354 void G1ParScanThreadState::trim_queue() { 4355 assert(_evac_cl != NULL, "not set"); 4356 assert(_evac_failure_cl != NULL, "not set"); 4357 assert(_partial_scan_cl != NULL, "not set"); 4358 4359 StarTask ref; 4360 do { 4361 // Drain the overflow stack first, so other threads can steal. 4362 while (refs()->pop_overflow(ref)) { 4363 deal_with_reference(ref); 4364 } 4365 4366 while (refs()->pop_local(ref)) { 4367 deal_with_reference(ref); 4368 } 4369 } while (!refs()->is_empty()); 4370 } 4371 4372 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, 4373 G1ParScanThreadState* par_scan_state) : 4374 _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()), 4375 _par_scan_state(par_scan_state), 4376 _worker_id(par_scan_state->queue_num()), 4377 _during_initial_mark(_g1->g1_policy()->during_initial_mark_pause()), 4378 _mark_in_progress(_g1->mark_in_progress()) { } 4379 4380 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4381 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object>::mark_object(oop obj) { 4382 #ifdef ASSERT 4383 HeapRegion* hr = _g1->heap_region_containing(obj); 4384 assert(hr != NULL, "sanity"); 4385 assert(!hr->in_collection_set(), "should not mark objects in the CSet"); 4386 #endif // ASSERT 4387 4388 // We know that the object is not moving so it's safe to read its size. 4389 _cm->grayRoot(obj, (size_t) obj->size(), _worker_id); 4390 } 4391 4392 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4393 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4394 ::mark_forwarded_object(oop from_obj, oop to_obj) { 4395 #ifdef ASSERT 4396 assert(from_obj->is_forwarded(), "from obj should be forwarded"); 4397 assert(from_obj->forwardee() == to_obj, "to obj should be the forwardee"); 4398 assert(from_obj != to_obj, "should not be self-forwarded"); 4399 4400 HeapRegion* from_hr = _g1->heap_region_containing(from_obj); 4401 assert(from_hr != NULL, "sanity"); 4402 assert(from_hr->in_collection_set(), "from obj should be in the CSet"); 4403 4404 HeapRegion* to_hr = _g1->heap_region_containing(to_obj); 4405 assert(to_hr != NULL, "sanity"); 4406 assert(!to_hr->in_collection_set(), "should not mark objects in the CSet"); 4407 #endif // ASSERT 4408 4409 // The object might be in the process of being copied by another 4410 // worker so we cannot trust that its to-space image is 4411 // well-formed. So we have to read its size from its from-space 4412 // image which we know should not be changing. 4413 _cm->grayRoot(to_obj, (size_t) from_obj->size(), _worker_id); 4414 } 4415 4416 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4417 oop G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4418 ::copy_to_survivor_space(oop old) { 4419 size_t word_sz = old->size(); 4420 HeapRegion* from_region = _g1->heap_region_containing_raw(old); 4421 // +1 to make the -1 indexes valid... 4422 int young_index = from_region->young_index_in_cset()+1; 4423 assert( (from_region->is_young() && young_index > 0) || 4424 (!from_region->is_young() && young_index == 0), "invariant" ); 4425 G1CollectorPolicy* g1p = _g1->g1_policy(); 4426 markOop m = old->mark(); 4427 int age = m->has_displaced_mark_helper() ? m->displaced_mark_helper()->age() 4428 : m->age(); 4429 GCAllocPurpose alloc_purpose = g1p->evacuation_destination(from_region, age, 4430 word_sz); 4431 HeapWord* obj_ptr = _par_scan_state->allocate(alloc_purpose, word_sz); 4432 oop obj = oop(obj_ptr); 4433 4434 if (obj_ptr == NULL) { 4435 // This will either forward-to-self, or detect that someone else has 4436 // installed a forwarding pointer. 4437 OopsInHeapRegionClosure* cl = _par_scan_state->evac_failure_closure(); 4438 return _g1->handle_evacuation_failure_par(cl, old); 4439 } 4440 4441 // We're going to allocate linearly, so might as well prefetch ahead. 4442 Prefetch::write(obj_ptr, PrefetchCopyIntervalInBytes); 4443 4444 oop forward_ptr = old->forward_to_atomic(obj); 4445 if (forward_ptr == NULL) { 4446 Copy::aligned_disjoint_words((HeapWord*) old, obj_ptr, word_sz); 4447 if (g1p->track_object_age(alloc_purpose)) { 4448 // We could simply do obj->incr_age(). However, this causes a 4449 // performance issue. obj->incr_age() will first check whether 4450 // the object has a displaced mark by checking its mark word; 4451 // getting the mark word from the new location of the object 4452 // stalls. So, given that we already have the mark word and we 4453 // are about to install it anyway, it's better to increase the 4454 // age on the mark word, when the object does not have a 4455 // displaced mark word. We're not expecting many objects to have 4456 // a displaced marked word, so that case is not optimized 4457 // further (it could be...) and we simply call obj->incr_age(). 4458 4459 if (m->has_displaced_mark_helper()) { 4460 // in this case, we have to install the mark word first, 4461 // otherwise obj looks to be forwarded (the old mark word, 4462 // which contains the forward pointer, was copied) 4463 obj->set_mark(m); 4464 obj->incr_age(); 4465 } else { 4466 m = m->incr_age(); 4467 obj->set_mark(m); 4468 } 4469 _par_scan_state->age_table()->add(obj, word_sz); 4470 } else { 4471 obj->set_mark(m); 4472 } 4473 4474 size_t* surv_young_words = _par_scan_state->surviving_young_words(); 4475 surv_young_words[young_index] += word_sz; 4476 4477 if (obj->is_objArray() && arrayOop(obj)->length() >= ParGCArrayScanChunk) { 4478 // We keep track of the next start index in the length field of 4479 // the to-space object. The actual length can be found in the 4480 // length field of the from-space object. 4481 arrayOop(obj)->set_length(0); 4482 oop* old_p = set_partial_array_mask(old); 4483 _par_scan_state->push_on_queue(old_p); 4484 } else { 4485 // No point in using the slower heap_region_containing() method, 4486 // given that we know obj is in the heap. 4487 _scanner.set_region(_g1->heap_region_containing_raw(obj)); 4488 obj->oop_iterate_backwards(&_scanner); 4489 } 4490 } else { 4491 _par_scan_state->undo_allocation(alloc_purpose, obj_ptr, word_sz); 4492 obj = forward_ptr; 4493 } 4494 return obj; 4495 } 4496 4497 template <bool do_gen_barrier, G1Barrier barrier, bool do_mark_object> 4498 template <class T> 4499 void G1ParCopyClosure<do_gen_barrier, barrier, do_mark_object> 4500 ::do_oop_work(T* p) { 4501 oop obj = oopDesc::load_decode_heap_oop(p); 4502 assert(barrier != G1BarrierRS || obj != NULL, 4503 "Precondition: G1BarrierRS implies obj is non-NULL"); 4504 4505 assert(_worker_id == _par_scan_state->queue_num(), "sanity"); 4506 4507 // here the null check is implicit in the cset_fast_test() test 4508 if (_g1->in_cset_fast_test(obj)) { 4509 oop forwardee; 4510 if (obj->is_forwarded()) { 4511 forwardee = obj->forwardee(); 4512 } else { 4513 forwardee = copy_to_survivor_space(obj); 4514 } 4515 assert(forwardee != NULL, "forwardee should not be NULL"); 4516 oopDesc::encode_store_heap_oop(p, forwardee); 4517 if (do_mark_object && forwardee != obj) { 4518 // If the object is self-forwarded we don't need to explicitly 4519 // mark it, the evacuation failure protocol will do so. 4520 mark_forwarded_object(obj, forwardee); 4521 } 4522 4523 // When scanning the RS, we only care about objs in CS. 4524 if (barrier == G1BarrierRS) { 4525 _par_scan_state->update_rs(_from, p, _worker_id); 4526 } 4527 } else { 4528 // The object is not in collection set. If we're a root scanning 4529 // closure during an initial mark pause (i.e. do_mark_object will 4530 // be true) then attempt to mark the object. 4531 if (do_mark_object && _g1->is_in_g1_reserved(obj)) { 4532 mark_object(obj); 4533 } 4534 } 4535 4536 if (barrier == G1BarrierEvac && obj != NULL) { 4537 _par_scan_state->update_rs(_from, p, _worker_id); 4538 } 4539 4540 if (do_gen_barrier && obj != NULL) { 4541 par_do_barrier(p); 4542 } 4543 } 4544 4545 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(oop* p); 4546 template void G1ParCopyClosure<false, G1BarrierEvac, false>::do_oop_work(narrowOop* p); 4547 4548 template <class T> void G1ParScanPartialArrayClosure::do_oop_nv(T* p) { 4549 assert(has_partial_array_mask(p), "invariant"); 4550 oop from_obj = clear_partial_array_mask(p); 4551 4552 assert(Universe::heap()->is_in_reserved(from_obj), "must be in heap."); 4553 assert(from_obj->is_objArray(), "must be obj array"); 4554 objArrayOop from_obj_array = objArrayOop(from_obj); 4555 // The from-space object contains the real length. 4556 int length = from_obj_array->length(); 4557 4558 assert(from_obj->is_forwarded(), "must be forwarded"); 4559 oop to_obj = from_obj->forwardee(); 4560 assert(from_obj != to_obj, "should not be chunking self-forwarded objects"); 4561 objArrayOop to_obj_array = objArrayOop(to_obj); 4562 // We keep track of the next start index in the length field of the 4563 // to-space object. 4564 int next_index = to_obj_array->length(); 4565 assert(0 <= next_index && next_index < length, 4566 err_msg("invariant, next index: %d, length: %d", next_index, length)); 4567 4568 int start = next_index; 4569 int end = length; 4570 int remainder = end - start; 4571 // We'll try not to push a range that's smaller than ParGCArrayScanChunk. 4572 if (remainder > 2 * ParGCArrayScanChunk) { 4573 end = start + ParGCArrayScanChunk; 4574 to_obj_array->set_length(end); 4575 // Push the remainder before we process the range in case another 4576 // worker has run out of things to do and can steal it. 4577 oop* from_obj_p = set_partial_array_mask(from_obj); 4578 _par_scan_state->push_on_queue(from_obj_p); 4579 } else { 4580 assert(length == end, "sanity"); 4581 // We'll process the final range for this object. Restore the length 4582 // so that the heap remains parsable in case of evacuation failure. 4583 to_obj_array->set_length(end); 4584 } 4585 _scanner.set_region(_g1->heap_region_containing_raw(to_obj)); 4586 // Process indexes [start,end). It will also process the header 4587 // along with the first chunk (i.e., the chunk with start == 0). 4588 // Note that at this point the length field of to_obj_array is not 4589 // correct given that we are using it to keep track of the next 4590 // start index. oop_iterate_range() (thankfully!) ignores the length 4591 // field and only relies on the start / end parameters. It does 4592 // however return the size of the object which will be incorrect. So 4593 // we have to ignore it even if we wanted to use it. 4594 to_obj_array->oop_iterate_range(&_scanner, start, end); 4595 } 4596 4597 class G1ParEvacuateFollowersClosure : public VoidClosure { 4598 protected: 4599 G1CollectedHeap* _g1h; 4600 G1ParScanThreadState* _par_scan_state; 4601 RefToScanQueueSet* _queues; 4602 ParallelTaskTerminator* _terminator; 4603 4604 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 4605 RefToScanQueueSet* queues() { return _queues; } 4606 ParallelTaskTerminator* terminator() { return _terminator; } 4607 4608 public: 4609 G1ParEvacuateFollowersClosure(G1CollectedHeap* g1h, 4610 G1ParScanThreadState* par_scan_state, 4611 RefToScanQueueSet* queues, 4612 ParallelTaskTerminator* terminator) 4613 : _g1h(g1h), _par_scan_state(par_scan_state), 4614 _queues(queues), _terminator(terminator) {} 4615 4616 void do_void(); 4617 4618 private: 4619 inline bool offer_termination(); 4620 }; 4621 4622 bool G1ParEvacuateFollowersClosure::offer_termination() { 4623 G1ParScanThreadState* const pss = par_scan_state(); 4624 pss->start_term_time(); 4625 const bool res = terminator()->offer_termination(); 4626 pss->end_term_time(); 4627 return res; 4628 } 4629 4630 void G1ParEvacuateFollowersClosure::do_void() { 4631 StarTask stolen_task; 4632 G1ParScanThreadState* const pss = par_scan_state(); 4633 pss->trim_queue(); 4634 4635 do { 4636 while (queues()->steal(pss->queue_num(), pss->hash_seed(), stolen_task)) { 4637 assert(pss->verify_task(stolen_task), "sanity"); 4638 if (stolen_task.is_narrow()) { 4639 pss->deal_with_reference((narrowOop*) stolen_task); 4640 } else { 4641 pss->deal_with_reference((oop*) stolen_task); 4642 } 4643 4644 // We've just processed a reference and we might have made 4645 // available new entries on the queues. So we have to make sure 4646 // we drain the queues as necessary. 4647 pss->trim_queue(); 4648 } 4649 } while (!offer_termination()); 4650 4651 pss->retire_alloc_buffers(); 4652 } 4653 4654 class G1ParTask : public AbstractGangTask { 4655 protected: 4656 G1CollectedHeap* _g1h; 4657 RefToScanQueueSet *_queues; 4658 ParallelTaskTerminator _terminator; 4659 uint _n_workers; 4660 4661 Mutex _stats_lock; 4662 Mutex* stats_lock() { return &_stats_lock; } 4663 4664 size_t getNCards() { 4665 return (_g1h->capacity() + G1BlockOffsetSharedArray::N_bytes - 1) 4666 / G1BlockOffsetSharedArray::N_bytes; 4667 } 4668 4669 public: 4670 G1ParTask(G1CollectedHeap* g1h, 4671 RefToScanQueueSet *task_queues) 4672 : AbstractGangTask("G1 collection"), 4673 _g1h(g1h), 4674 _queues(task_queues), 4675 _terminator(0, _queues), 4676 _stats_lock(Mutex::leaf, "parallel G1 stats lock", true) 4677 {} 4678 4679 RefToScanQueueSet* queues() { return _queues; } 4680 4681 RefToScanQueue *work_queue(int i) { 4682 return queues()->queue(i); 4683 } 4684 4685 ParallelTaskTerminator* terminator() { return &_terminator; } 4686 4687 virtual void set_for_termination(int active_workers) { 4688 // This task calls set_n_termination() in par_non_clean_card_iterate_work() 4689 // in the young space (_par_seq_tasks) in the G1 heap 4690 // for SequentialSubTasksDone. 4691 // This task also uses SubTasksDone in SharedHeap and G1CollectedHeap 4692 // both of which need setting by set_n_termination(). 4693 _g1h->SharedHeap::set_n_termination(active_workers); 4694 _g1h->set_n_termination(active_workers); 4695 terminator()->reset_for_reuse(active_workers); 4696 _n_workers = active_workers; 4697 } 4698 4699 void work(uint worker_id) { 4700 if (worker_id >= _n_workers) return; // no work needed this round 4701 4702 double start_time_ms = os::elapsedTime() * 1000.0; 4703 _g1h->g1_policy()->phase_times()->record_gc_worker_start_time(worker_id, start_time_ms); 4704 4705 { 4706 ResourceMark rm; 4707 HandleMark hm; 4708 4709 ReferenceProcessor* rp = _g1h->ref_processor_stw(); 4710 4711 G1ParScanThreadState pss(_g1h, worker_id); 4712 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, rp); 4713 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, rp); 4714 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, rp); 4715 4716 pss.set_evac_closure(&scan_evac_cl); 4717 pss.set_evac_failure_closure(&evac_failure_cl); 4718 pss.set_partial_scan_closure(&partial_scan_cl); 4719 4720 G1ParScanExtRootClosure only_scan_root_cl(_g1h, &pss, rp); 4721 G1ParScanPermClosure only_scan_perm_cl(_g1h, &pss, rp); 4722 4723 G1ParScanAndMarkExtRootClosure scan_mark_root_cl(_g1h, &pss, rp); 4724 G1ParScanAndMarkPermClosure scan_mark_perm_cl(_g1h, &pss, rp); 4725 4726 OopClosure* scan_root_cl = &only_scan_root_cl; 4727 OopsInHeapRegionClosure* scan_perm_cl = &only_scan_perm_cl; 4728 4729 if (_g1h->g1_policy()->during_initial_mark_pause()) { 4730 // We also need to mark copied objects. 4731 scan_root_cl = &scan_mark_root_cl; 4732 scan_perm_cl = &scan_mark_perm_cl; 4733 } 4734 4735 G1ParPushHeapRSClosure push_heap_rs_cl(_g1h, &pss); 4736 4737 pss.start_strong_roots(); 4738 _g1h->g1_process_strong_roots(/* not collecting perm */ false, 4739 SharedHeap::SO_AllClasses, 4740 scan_root_cl, 4741 &push_heap_rs_cl, 4742 scan_perm_cl, 4743 worker_id); 4744 pss.end_strong_roots(); 4745 4746 { 4747 double start = os::elapsedTime(); 4748 G1ParEvacuateFollowersClosure evac(_g1h, &pss, _queues, &_terminator); 4749 evac.do_void(); 4750 double elapsed_ms = (os::elapsedTime()-start)*1000.0; 4751 double term_ms = pss.term_time()*1000.0; 4752 _g1h->g1_policy()->phase_times()->add_obj_copy_time(worker_id, elapsed_ms-term_ms); 4753 _g1h->g1_policy()->phase_times()->record_termination(worker_id, term_ms, pss.term_attempts()); 4754 } 4755 _g1h->g1_policy()->record_thread_age_table(pss.age_table()); 4756 _g1h->update_surviving_young_words(pss.surviving_young_words()+1); 4757 4758 if (ParallelGCVerbose) { 4759 MutexLocker x(stats_lock()); 4760 pss.print_termination_stats(worker_id); 4761 } 4762 4763 assert(pss.refs()->is_empty(), "should be empty"); 4764 4765 // Close the inner scope so that the ResourceMark and HandleMark 4766 // destructors are executed here and are included as part of the 4767 // "GC Worker Time". 4768 } 4769 4770 double end_time_ms = os::elapsedTime() * 1000.0; 4771 _g1h->g1_policy()->phase_times()->record_gc_worker_end_time(worker_id, end_time_ms); 4772 } 4773 }; 4774 4775 // *** Common G1 Evacuation Stuff 4776 4777 // Closures that support the filtering of CodeBlobs scanned during 4778 // external root scanning. 4779 4780 // Closure applied to reference fields in code blobs (specifically nmethods) 4781 // to determine whether an nmethod contains references that point into 4782 // the collection set. Used as a predicate when walking code roots so 4783 // that only nmethods that point into the collection set are added to the 4784 // 'marked' list. 4785 4786 class G1FilteredCodeBlobToOopClosure : public CodeBlobToOopClosure { 4787 4788 class G1PointsIntoCSOopClosure : public OopClosure { 4789 G1CollectedHeap* _g1; 4790 bool _points_into_cs; 4791 public: 4792 G1PointsIntoCSOopClosure(G1CollectedHeap* g1) : 4793 _g1(g1), _points_into_cs(false) { } 4794 4795 bool points_into_cs() const { return _points_into_cs; } 4796 4797 template <class T> 4798 void do_oop_nv(T* p) { 4799 if (!_points_into_cs) { 4800 T heap_oop = oopDesc::load_heap_oop(p); 4801 if (!oopDesc::is_null(heap_oop) && 4802 _g1->in_cset_fast_test(oopDesc::decode_heap_oop_not_null(heap_oop))) { 4803 _points_into_cs = true; 4804 } 4805 } 4806 } 4807 4808 virtual void do_oop(oop* p) { do_oop_nv(p); } 4809 virtual void do_oop(narrowOop* p) { do_oop_nv(p); } 4810 }; 4811 4812 G1CollectedHeap* _g1; 4813 4814 public: 4815 G1FilteredCodeBlobToOopClosure(G1CollectedHeap* g1, OopClosure* cl) : 4816 CodeBlobToOopClosure(cl, true), _g1(g1) { } 4817 4818 virtual void do_code_blob(CodeBlob* cb) { 4819 nmethod* nm = cb->as_nmethod_or_null(); 4820 if (nm != NULL && !(nm->test_oops_do_mark())) { 4821 G1PointsIntoCSOopClosure predicate_cl(_g1); 4822 nm->oops_do(&predicate_cl); 4823 4824 if (predicate_cl.points_into_cs()) { 4825 // At least one of the reference fields or the oop relocations 4826 // in the nmethod points into the collection set. We have to 4827 // 'mark' this nmethod. 4828 // Note: Revisit the following if CodeBlobToOopClosure::do_code_blob() 4829 // or MarkingCodeBlobClosure::do_code_blob() change. 4830 if (!nm->test_set_oops_do_mark()) { 4831 do_newly_marked_nmethod(nm); 4832 } 4833 } 4834 } 4835 } 4836 }; 4837 4838 // This method is run in a GC worker. 4839 4840 void 4841 G1CollectedHeap:: 4842 g1_process_strong_roots(bool collecting_perm_gen, 4843 ScanningOption so, 4844 OopClosure* scan_non_heap_roots, 4845 OopsInHeapRegionClosure* scan_rs, 4846 OopsInGenClosure* scan_perm, 4847 int worker_i) { 4848 4849 // First scan the strong roots, including the perm gen. 4850 double ext_roots_start = os::elapsedTime(); 4851 double closure_app_time_sec = 0.0; 4852 4853 BufferingOopClosure buf_scan_non_heap_roots(scan_non_heap_roots); 4854 BufferingOopsInGenClosure buf_scan_perm(scan_perm); 4855 buf_scan_perm.set_generation(perm_gen()); 4856 4857 // Walk the code cache w/o buffering, because StarTask cannot handle 4858 // unaligned oop locations. 4859 G1FilteredCodeBlobToOopClosure eager_scan_code_roots(this, scan_non_heap_roots); 4860 4861 process_strong_roots(false, // no scoping; this is parallel code 4862 collecting_perm_gen, so, 4863 &buf_scan_non_heap_roots, 4864 &eager_scan_code_roots, 4865 &buf_scan_perm); 4866 4867 // Now the CM ref_processor roots. 4868 if (!_process_strong_tasks->is_task_claimed(G1H_PS_refProcessor_oops_do)) { 4869 // We need to treat the discovered reference lists of the 4870 // concurrent mark ref processor as roots and keep entries 4871 // (which are added by the marking threads) on them live 4872 // until they can be processed at the end of marking. 4873 ref_processor_cm()->weak_oops_do(&buf_scan_non_heap_roots); 4874 } 4875 4876 // Finish up any enqueued closure apps (attributed as object copy time). 4877 buf_scan_non_heap_roots.done(); 4878 buf_scan_perm.done(); 4879 4880 double obj_copy_time_sec = buf_scan_perm.closure_app_seconds() + 4881 buf_scan_non_heap_roots.closure_app_seconds(); 4882 g1_policy()->phase_times()->record_obj_copy_time(worker_i, obj_copy_time_sec * 1000.0); 4883 4884 double ext_root_time_ms = 4885 ((os::elapsedTime() - ext_roots_start) - obj_copy_time_sec) * 1000.0; 4886 4887 g1_policy()->phase_times()->record_ext_root_scan_time(worker_i, ext_root_time_ms); 4888 4889 // During conc marking we have to filter the per-thread SATB buffers 4890 // to make sure we remove any oops into the CSet (which will show up 4891 // as implicitly live). 4892 double satb_filtering_ms = 0.0; 4893 if (!_process_strong_tasks->is_task_claimed(G1H_PS_filter_satb_buffers)) { 4894 if (mark_in_progress()) { 4895 double satb_filter_start = os::elapsedTime(); 4896 4897 JavaThread::satb_mark_queue_set().filter_thread_buffers(); 4898 4899 satb_filtering_ms = (os::elapsedTime() - satb_filter_start) * 1000.0; 4900 } 4901 } 4902 g1_policy()->phase_times()->record_satb_filtering_time(worker_i, satb_filtering_ms); 4903 4904 // Now scan the complement of the collection set. 4905 if (scan_rs != NULL) { 4906 g1_rem_set()->oops_into_collection_set_do(scan_rs, worker_i); 4907 } 4908 4909 _process_strong_tasks->all_tasks_completed(); 4910 } 4911 4912 void 4913 G1CollectedHeap::g1_process_weak_roots(OopClosure* root_closure, 4914 OopClosure* non_root_closure) { 4915 CodeBlobToOopClosure roots_in_blobs(root_closure, /*do_marking=*/ false); 4916 SharedHeap::process_weak_roots(root_closure, &roots_in_blobs, non_root_closure); 4917 } 4918 4919 // Weak Reference Processing support 4920 4921 // An always "is_alive" closure that is used to preserve referents. 4922 // If the object is non-null then it's alive. Used in the preservation 4923 // of referent objects that are pointed to by reference objects 4924 // discovered by the CM ref processor. 4925 class G1AlwaysAliveClosure: public BoolObjectClosure { 4926 G1CollectedHeap* _g1; 4927 public: 4928 G1AlwaysAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4929 void do_object(oop p) { assert(false, "Do not call."); } 4930 bool do_object_b(oop p) { 4931 if (p != NULL) { 4932 return true; 4933 } 4934 return false; 4935 } 4936 }; 4937 4938 bool G1STWIsAliveClosure::do_object_b(oop p) { 4939 // An object is reachable if it is outside the collection set, 4940 // or is inside and copied. 4941 return !_g1->obj_in_cs(p) || p->is_forwarded(); 4942 } 4943 4944 // Non Copying Keep Alive closure 4945 class G1KeepAliveClosure: public OopClosure { 4946 G1CollectedHeap* _g1; 4947 public: 4948 G1KeepAliveClosure(G1CollectedHeap* g1) : _g1(g1) {} 4949 void do_oop(narrowOop* p) { guarantee(false, "Not needed"); } 4950 void do_oop( oop* p) { 4951 oop obj = *p; 4952 4953 if (_g1->obj_in_cs(obj)) { 4954 assert( obj->is_forwarded(), "invariant" ); 4955 *p = obj->forwardee(); 4956 } 4957 } 4958 }; 4959 4960 // Copying Keep Alive closure - can be called from both 4961 // serial and parallel code as long as different worker 4962 // threads utilize different G1ParScanThreadState instances 4963 // and different queues. 4964 4965 class G1CopyingKeepAliveClosure: public OopClosure { 4966 G1CollectedHeap* _g1h; 4967 OopClosure* _copy_non_heap_obj_cl; 4968 OopsInHeapRegionClosure* _copy_perm_obj_cl; 4969 G1ParScanThreadState* _par_scan_state; 4970 4971 public: 4972 G1CopyingKeepAliveClosure(G1CollectedHeap* g1h, 4973 OopClosure* non_heap_obj_cl, 4974 OopsInHeapRegionClosure* perm_obj_cl, 4975 G1ParScanThreadState* pss): 4976 _g1h(g1h), 4977 _copy_non_heap_obj_cl(non_heap_obj_cl), 4978 _copy_perm_obj_cl(perm_obj_cl), 4979 _par_scan_state(pss) 4980 {} 4981 4982 virtual void do_oop(narrowOop* p) { do_oop_work(p); } 4983 virtual void do_oop( oop* p) { do_oop_work(p); } 4984 4985 template <class T> void do_oop_work(T* p) { 4986 oop obj = oopDesc::load_decode_heap_oop(p); 4987 4988 if (_g1h->obj_in_cs(obj)) { 4989 // If the referent object has been forwarded (either copied 4990 // to a new location or to itself in the event of an 4991 // evacuation failure) then we need to update the reference 4992 // field and, if both reference and referent are in the G1 4993 // heap, update the RSet for the referent. 4994 // 4995 // If the referent has not been forwarded then we have to keep 4996 // it alive by policy. Therefore we have copy the referent. 4997 // 4998 // If the reference field is in the G1 heap then we can push 4999 // on the PSS queue. When the queue is drained (after each 5000 // phase of reference processing) the object and it's followers 5001 // will be copied, the reference field set to point to the 5002 // new location, and the RSet updated. Otherwise we need to 5003 // use the the non-heap or perm closures directly to copy 5004 // the refernt object and update the pointer, while avoiding 5005 // updating the RSet. 5006 5007 if (_g1h->is_in_g1_reserved(p)) { 5008 _par_scan_state->push_on_queue(p); 5009 } else { 5010 // The reference field is not in the G1 heap. 5011 if (_g1h->perm_gen()->is_in(p)) { 5012 _copy_perm_obj_cl->do_oop(p); 5013 } else { 5014 _copy_non_heap_obj_cl->do_oop(p); 5015 } 5016 } 5017 } 5018 } 5019 }; 5020 5021 // Serial drain queue closure. Called as the 'complete_gc' 5022 // closure for each discovered list in some of the 5023 // reference processing phases. 5024 5025 class G1STWDrainQueueClosure: public VoidClosure { 5026 protected: 5027 G1CollectedHeap* _g1h; 5028 G1ParScanThreadState* _par_scan_state; 5029 5030 G1ParScanThreadState* par_scan_state() { return _par_scan_state; } 5031 5032 public: 5033 G1STWDrainQueueClosure(G1CollectedHeap* g1h, G1ParScanThreadState* pss) : 5034 _g1h(g1h), 5035 _par_scan_state(pss) 5036 { } 5037 5038 void do_void() { 5039 G1ParScanThreadState* const pss = par_scan_state(); 5040 pss->trim_queue(); 5041 } 5042 }; 5043 5044 // Parallel Reference Processing closures 5045 5046 // Implementation of AbstractRefProcTaskExecutor for parallel reference 5047 // processing during G1 evacuation pauses. 5048 5049 class G1STWRefProcTaskExecutor: public AbstractRefProcTaskExecutor { 5050 private: 5051 G1CollectedHeap* _g1h; 5052 RefToScanQueueSet* _queues; 5053 FlexibleWorkGang* _workers; 5054 int _active_workers; 5055 5056 public: 5057 G1STWRefProcTaskExecutor(G1CollectedHeap* g1h, 5058 FlexibleWorkGang* workers, 5059 RefToScanQueueSet *task_queues, 5060 int n_workers) : 5061 _g1h(g1h), 5062 _queues(task_queues), 5063 _workers(workers), 5064 _active_workers(n_workers) 5065 { 5066 assert(n_workers > 0, "shouldn't call this otherwise"); 5067 } 5068 5069 // Executes the given task using concurrent marking worker threads. 5070 virtual void execute(ProcessTask& task); 5071 virtual void execute(EnqueueTask& task); 5072 }; 5073 5074 // Gang task for possibly parallel reference processing 5075 5076 class G1STWRefProcTaskProxy: public AbstractGangTask { 5077 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; 5078 ProcessTask& _proc_task; 5079 G1CollectedHeap* _g1h; 5080 RefToScanQueueSet *_task_queues; 5081 ParallelTaskTerminator* _terminator; 5082 5083 public: 5084 G1STWRefProcTaskProxy(ProcessTask& proc_task, 5085 G1CollectedHeap* g1h, 5086 RefToScanQueueSet *task_queues, 5087 ParallelTaskTerminator* terminator) : 5088 AbstractGangTask("Process reference objects in parallel"), 5089 _proc_task(proc_task), 5090 _g1h(g1h), 5091 _task_queues(task_queues), 5092 _terminator(terminator) 5093 {} 5094 5095 virtual void work(uint worker_id) { 5096 // The reference processing task executed by a single worker. 5097 ResourceMark rm; 5098 HandleMark hm; 5099 5100 G1STWIsAliveClosure is_alive(_g1h); 5101 5102 G1ParScanThreadState pss(_g1h, worker_id); 5103 5104 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5105 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5106 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5107 5108 pss.set_evac_closure(&scan_evac_cl); 5109 pss.set_evac_failure_closure(&evac_failure_cl); 5110 pss.set_partial_scan_closure(&partial_scan_cl); 5111 5112 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5113 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5114 5115 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5116 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5117 5118 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5119 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5120 5121 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5122 // We also need to mark copied objects. 5123 copy_non_heap_cl = ©_mark_non_heap_cl; 5124 copy_perm_cl = ©_mark_perm_cl; 5125 } 5126 5127 // Keep alive closure. 5128 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5129 5130 // Complete GC closure 5131 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _task_queues, _terminator); 5132 5133 // Call the reference processing task's work routine. 5134 _proc_task.work(worker_id, is_alive, keep_alive, drain_queue); 5135 5136 // Note we cannot assert that the refs array is empty here as not all 5137 // of the processing tasks (specifically phase2 - pp2_work) execute 5138 // the complete_gc closure (which ordinarily would drain the queue) so 5139 // the queue may not be empty. 5140 } 5141 }; 5142 5143 // Driver routine for parallel reference processing. 5144 // Creates an instance of the ref processing gang 5145 // task and has the worker threads execute it. 5146 void G1STWRefProcTaskExecutor::execute(ProcessTask& proc_task) { 5147 assert(_workers != NULL, "Need parallel worker threads."); 5148 5149 ParallelTaskTerminator terminator(_active_workers, _queues); 5150 G1STWRefProcTaskProxy proc_task_proxy(proc_task, _g1h, _queues, &terminator); 5151 5152 _g1h->set_par_threads(_active_workers); 5153 _workers->run_task(&proc_task_proxy); 5154 _g1h->set_par_threads(0); 5155 } 5156 5157 // Gang task for parallel reference enqueueing. 5158 5159 class G1STWRefEnqueueTaskProxy: public AbstractGangTask { 5160 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; 5161 EnqueueTask& _enq_task; 5162 5163 public: 5164 G1STWRefEnqueueTaskProxy(EnqueueTask& enq_task) : 5165 AbstractGangTask("Enqueue reference objects in parallel"), 5166 _enq_task(enq_task) 5167 { } 5168 5169 virtual void work(uint worker_id) { 5170 _enq_task.work(worker_id); 5171 } 5172 }; 5173 5174 // Driver routine for parallel reference enqueing. 5175 // Creates an instance of the ref enqueueing gang 5176 // task and has the worker threads execute it. 5177 5178 void G1STWRefProcTaskExecutor::execute(EnqueueTask& enq_task) { 5179 assert(_workers != NULL, "Need parallel worker threads."); 5180 5181 G1STWRefEnqueueTaskProxy enq_task_proxy(enq_task); 5182 5183 _g1h->set_par_threads(_active_workers); 5184 _workers->run_task(&enq_task_proxy); 5185 _g1h->set_par_threads(0); 5186 } 5187 5188 // End of weak reference support closures 5189 5190 // Abstract task used to preserve (i.e. copy) any referent objects 5191 // that are in the collection set and are pointed to by reference 5192 // objects discovered by the CM ref processor. 5193 5194 class G1ParPreserveCMReferentsTask: public AbstractGangTask { 5195 protected: 5196 G1CollectedHeap* _g1h; 5197 RefToScanQueueSet *_queues; 5198 ParallelTaskTerminator _terminator; 5199 uint _n_workers; 5200 5201 public: 5202 G1ParPreserveCMReferentsTask(G1CollectedHeap* g1h,int workers, RefToScanQueueSet *task_queues) : 5203 AbstractGangTask("ParPreserveCMReferents"), 5204 _g1h(g1h), 5205 _queues(task_queues), 5206 _terminator(workers, _queues), 5207 _n_workers(workers) 5208 { } 5209 5210 void work(uint worker_id) { 5211 ResourceMark rm; 5212 HandleMark hm; 5213 5214 G1ParScanThreadState pss(_g1h, worker_id); 5215 G1ParScanHeapEvacClosure scan_evac_cl(_g1h, &pss, NULL); 5216 G1ParScanHeapEvacFailureClosure evac_failure_cl(_g1h, &pss, NULL); 5217 G1ParScanPartialArrayClosure partial_scan_cl(_g1h, &pss, NULL); 5218 5219 pss.set_evac_closure(&scan_evac_cl); 5220 pss.set_evac_failure_closure(&evac_failure_cl); 5221 pss.set_partial_scan_closure(&partial_scan_cl); 5222 5223 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5224 5225 5226 G1ParScanExtRootClosure only_copy_non_heap_cl(_g1h, &pss, NULL); 5227 G1ParScanPermClosure only_copy_perm_cl(_g1h, &pss, NULL); 5228 5229 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(_g1h, &pss, NULL); 5230 G1ParScanAndMarkPermClosure copy_mark_perm_cl(_g1h, &pss, NULL); 5231 5232 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5233 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5234 5235 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5236 // We also need to mark copied objects. 5237 copy_non_heap_cl = ©_mark_non_heap_cl; 5238 copy_perm_cl = ©_mark_perm_cl; 5239 } 5240 5241 // Is alive closure 5242 G1AlwaysAliveClosure always_alive(_g1h); 5243 5244 // Copying keep alive closure. Applied to referent objects that need 5245 // to be copied. 5246 G1CopyingKeepAliveClosure keep_alive(_g1h, copy_non_heap_cl, copy_perm_cl, &pss); 5247 5248 ReferenceProcessor* rp = _g1h->ref_processor_cm(); 5249 5250 uint limit = ReferenceProcessor::number_of_subclasses_of_ref() * rp->max_num_q(); 5251 uint stride = MIN2(MAX2(_n_workers, 1U), limit); 5252 5253 // limit is set using max_num_q() - which was set using ParallelGCThreads. 5254 // So this must be true - but assert just in case someone decides to 5255 // change the worker ids. 5256 assert(0 <= worker_id && worker_id < limit, "sanity"); 5257 assert(!rp->discovery_is_atomic(), "check this code"); 5258 5259 // Select discovered lists [i, i+stride, i+2*stride,...,limit) 5260 for (uint idx = worker_id; idx < limit; idx += stride) { 5261 DiscoveredList& ref_list = rp->discovered_refs()[idx]; 5262 5263 DiscoveredListIterator iter(ref_list, &keep_alive, &always_alive); 5264 while (iter.has_next()) { 5265 // Since discovery is not atomic for the CM ref processor, we 5266 // can see some null referent objects. 5267 iter.load_ptrs(DEBUG_ONLY(true)); 5268 oop ref = iter.obj(); 5269 5270 // This will filter nulls. 5271 if (iter.is_referent_alive()) { 5272 iter.make_referent_alive(); 5273 } 5274 iter.move_to_next(); 5275 } 5276 } 5277 5278 // Drain the queue - which may cause stealing 5279 G1ParEvacuateFollowersClosure drain_queue(_g1h, &pss, _queues, &_terminator); 5280 drain_queue.do_void(); 5281 // Allocation buffers were retired at the end of G1ParEvacuateFollowersClosure 5282 assert(pss.refs()->is_empty(), "should be"); 5283 } 5284 }; 5285 5286 // Weak Reference processing during an evacuation pause (part 1). 5287 void G1CollectedHeap::process_discovered_references() { 5288 double ref_proc_start = os::elapsedTime(); 5289 5290 ReferenceProcessor* rp = _ref_processor_stw; 5291 assert(rp->discovery_enabled(), "should have been enabled"); 5292 5293 // Any reference objects, in the collection set, that were 'discovered' 5294 // by the CM ref processor should have already been copied (either by 5295 // applying the external root copy closure to the discovered lists, or 5296 // by following an RSet entry). 5297 // 5298 // But some of the referents, that are in the collection set, that these 5299 // reference objects point to may not have been copied: the STW ref 5300 // processor would have seen that the reference object had already 5301 // been 'discovered' and would have skipped discovering the reference, 5302 // but would not have treated the reference object as a regular oop. 5303 // As a reult the copy closure would not have been applied to the 5304 // referent object. 5305 // 5306 // We need to explicitly copy these referent objects - the references 5307 // will be processed at the end of remarking. 5308 // 5309 // We also need to do this copying before we process the reference 5310 // objects discovered by the STW ref processor in case one of these 5311 // referents points to another object which is also referenced by an 5312 // object discovered by the STW ref processor. 5313 5314 uint active_workers = (G1CollectedHeap::use_parallel_gc_threads() ? 5315 workers()->active_workers() : 1); 5316 5317 assert(!G1CollectedHeap::use_parallel_gc_threads() || 5318 active_workers == workers()->active_workers(), 5319 "Need to reset active_workers"); 5320 5321 set_par_threads(active_workers); 5322 G1ParPreserveCMReferentsTask keep_cm_referents(this, active_workers, _task_queues); 5323 5324 if (G1CollectedHeap::use_parallel_gc_threads()) { 5325 workers()->run_task(&keep_cm_referents); 5326 } else { 5327 keep_cm_referents.work(0); 5328 } 5329 5330 set_par_threads(0); 5331 5332 // Closure to test whether a referent is alive. 5333 G1STWIsAliveClosure is_alive(this); 5334 5335 // Even when parallel reference processing is enabled, the processing 5336 // of JNI refs is serial and performed serially by the current thread 5337 // rather than by a worker. The following PSS will be used for processing 5338 // JNI refs. 5339 5340 // Use only a single queue for this PSS. 5341 G1ParScanThreadState pss(this, 0); 5342 5343 // We do not embed a reference processor in the copying/scanning 5344 // closures while we're actually processing the discovered 5345 // reference objects. 5346 G1ParScanHeapEvacClosure scan_evac_cl(this, &pss, NULL); 5347 G1ParScanHeapEvacFailureClosure evac_failure_cl(this, &pss, NULL); 5348 G1ParScanPartialArrayClosure partial_scan_cl(this, &pss, NULL); 5349 5350 pss.set_evac_closure(&scan_evac_cl); 5351 pss.set_evac_failure_closure(&evac_failure_cl); 5352 pss.set_partial_scan_closure(&partial_scan_cl); 5353 5354 assert(pss.refs()->is_empty(), "pre-condition"); 5355 5356 G1ParScanExtRootClosure only_copy_non_heap_cl(this, &pss, NULL); 5357 G1ParScanPermClosure only_copy_perm_cl(this, &pss, NULL); 5358 5359 G1ParScanAndMarkExtRootClosure copy_mark_non_heap_cl(this, &pss, NULL); 5360 G1ParScanAndMarkPermClosure copy_mark_perm_cl(this, &pss, NULL); 5361 5362 OopClosure* copy_non_heap_cl = &only_copy_non_heap_cl; 5363 OopsInHeapRegionClosure* copy_perm_cl = &only_copy_perm_cl; 5364 5365 if (_g1h->g1_policy()->during_initial_mark_pause()) { 5366 // We also need to mark copied objects. 5367 copy_non_heap_cl = ©_mark_non_heap_cl; 5368 copy_perm_cl = ©_mark_perm_cl; 5369 } 5370 5371 // Keep alive closure. 5372 G1CopyingKeepAliveClosure keep_alive(this, copy_non_heap_cl, copy_perm_cl, &pss); 5373 5374 // Serial Complete GC closure 5375 G1STWDrainQueueClosure drain_queue(this, &pss); 5376 5377 // Setup the soft refs policy... 5378 rp->setup_policy(false); 5379 5380 if (!rp->processing_is_mt()) { 5381 // Serial reference processing... 5382 rp->process_discovered_references(&is_alive, 5383 &keep_alive, 5384 &drain_queue, 5385 NULL); 5386 } else { 5387 // Parallel reference processing 5388 assert(rp->num_q() == active_workers, "sanity"); 5389 assert(active_workers <= rp->max_num_q(), "sanity"); 5390 5391 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5392 rp->process_discovered_references(&is_alive, &keep_alive, &drain_queue, &par_task_executor); 5393 } 5394 5395 // We have completed copying any necessary live referent objects 5396 // (that were not copied during the actual pause) so we can 5397 // retire any active alloc buffers 5398 pss.retire_alloc_buffers(); 5399 assert(pss.refs()->is_empty(), "both queue and overflow should be empty"); 5400 5401 double ref_proc_time = os::elapsedTime() - ref_proc_start; 5402 g1_policy()->phase_times()->record_ref_proc_time(ref_proc_time * 1000.0); 5403 } 5404 5405 // Weak Reference processing during an evacuation pause (part 2). 5406 void G1CollectedHeap::enqueue_discovered_references() { 5407 double ref_enq_start = os::elapsedTime(); 5408 5409 ReferenceProcessor* rp = _ref_processor_stw; 5410 assert(!rp->discovery_enabled(), "should have been disabled as part of processing"); 5411 5412 // Now enqueue any remaining on the discovered lists on to 5413 // the pending list. 5414 if (!rp->processing_is_mt()) { 5415 // Serial reference processing... 5416 rp->enqueue_discovered_references(); 5417 } else { 5418 // Parallel reference enqueuing 5419 5420 uint active_workers = (ParallelGCThreads > 0 ? workers()->active_workers() : 1); 5421 assert(active_workers == workers()->active_workers(), 5422 "Need to reset active_workers"); 5423 assert(rp->num_q() == active_workers, "sanity"); 5424 assert(active_workers <= rp->max_num_q(), "sanity"); 5425 5426 G1STWRefProcTaskExecutor par_task_executor(this, workers(), _task_queues, active_workers); 5427 rp->enqueue_discovered_references(&par_task_executor); 5428 } 5429 5430 rp->verify_no_references_recorded(); 5431 assert(!rp->discovery_enabled(), "should have been disabled"); 5432 5433 // FIXME 5434 // CM's reference processing also cleans up the string and symbol tables. 5435 // Should we do that here also? We could, but it is a serial operation 5436 // and could signicantly increase the pause time. 5437 5438 double ref_enq_time = os::elapsedTime() - ref_enq_start; 5439 g1_policy()->phase_times()->record_ref_enq_time(ref_enq_time * 1000.0); 5440 } 5441 5442 void G1CollectedHeap::evacuate_collection_set() { 5443 _expand_heap_after_alloc_failure = true; 5444 set_evacuation_failed(false); 5445 5446 g1_rem_set()->prepare_for_oops_into_collection_set_do(); 5447 concurrent_g1_refine()->set_use_cache(false); 5448 concurrent_g1_refine()->clear_hot_cache_claimed_index(); 5449 5450 uint n_workers; 5451 if (G1CollectedHeap::use_parallel_gc_threads()) { 5452 n_workers = 5453 AdaptiveSizePolicy::calc_active_workers(workers()->total_workers(), 5454 workers()->active_workers(), 5455 Threads::number_of_non_daemon_threads()); 5456 assert(UseDynamicNumberOfGCThreads || 5457 n_workers == workers()->total_workers(), 5458 "If not dynamic should be using all the workers"); 5459 workers()->set_active_workers(n_workers); 5460 set_par_threads(n_workers); 5461 } else { 5462 assert(n_par_threads() == 0, 5463 "Should be the original non-parallel value"); 5464 n_workers = 1; 5465 } 5466 5467 G1ParTask g1_par_task(this, _task_queues); 5468 5469 init_for_evac_failure(NULL); 5470 5471 rem_set()->prepare_for_younger_refs_iterate(true); 5472 5473 assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty"); 5474 double start_par_time_sec = os::elapsedTime(); 5475 double end_par_time_sec; 5476 5477 { 5478 StrongRootsScope srs(this); 5479 5480 if (G1CollectedHeap::use_parallel_gc_threads()) { 5481 // The individual threads will set their evac-failure closures. 5482 if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr(); 5483 // These tasks use ShareHeap::_process_strong_tasks 5484 assert(UseDynamicNumberOfGCThreads || 5485 workers()->active_workers() == workers()->total_workers(), 5486 "If not dynamic should be using all the workers"); 5487 workers()->run_task(&g1_par_task); 5488 } else { 5489 g1_par_task.set_for_termination(n_workers); 5490 g1_par_task.work(0); 5491 } 5492 end_par_time_sec = os::elapsedTime(); 5493 5494 // Closing the inner scope will execute the destructor 5495 // for the StrongRootsScope object. We record the current 5496 // elapsed time before closing the scope so that time 5497 // taken for the SRS destructor is NOT included in the 5498 // reported parallel time. 5499 } 5500 5501 double par_time_ms = (end_par_time_sec - start_par_time_sec) * 1000.0; 5502 g1_policy()->phase_times()->record_par_time(par_time_ms); 5503 5504 double code_root_fixup_time_ms = 5505 (os::elapsedTime() - end_par_time_sec) * 1000.0; 5506 g1_policy()->phase_times()->record_code_root_fixup_time(code_root_fixup_time_ms); 5507 5508 set_par_threads(0); 5509 5510 // Process any discovered reference objects - we have 5511 // to do this _before_ we retire the GC alloc regions 5512 // as we may have to copy some 'reachable' referent 5513 // objects (and their reachable sub-graphs) that were 5514 // not copied during the pause. 5515 process_discovered_references(); 5516 5517 // Weak root processing. 5518 // Note: when JSR 292 is enabled and code blobs can contain 5519 // non-perm oops then we will need to process the code blobs 5520 // here too. 5521 { 5522 G1STWIsAliveClosure is_alive(this); 5523 G1KeepAliveClosure keep_alive(this); 5524 JNIHandles::weak_oops_do(&is_alive, &keep_alive); 5525 } 5526 5527 release_gc_alloc_regions(); 5528 g1_rem_set()->cleanup_after_oops_into_collection_set_do(); 5529 5530 concurrent_g1_refine()->clear_hot_cache(); 5531 concurrent_g1_refine()->set_use_cache(true); 5532 5533 finalize_for_evac_failure(); 5534 5535 if (evacuation_failed()) { 5536 remove_self_forwarding_pointers(); 5537 if (G1Log::finer()) { 5538 gclog_or_tty->print(" (to-space exhausted)"); 5539 } else if (G1Log::fine()) { 5540 gclog_or_tty->print("--"); 5541 } 5542 } 5543 5544 // Enqueue any remaining references remaining on the STW 5545 // reference processor's discovered lists. We need to do 5546 // this after the card table is cleaned (and verified) as 5547 // the act of enqueuing entries on to the pending list 5548 // will log these updates (and dirty their associated 5549 // cards). We need these updates logged to update any 5550 // RSets. 5551 enqueue_discovered_references(); 5552 5553 if (G1DeferredRSUpdate) { 5554 RedirtyLoggedCardTableEntryFastClosure redirty; 5555 dirty_card_queue_set().set_closure(&redirty); 5556 dirty_card_queue_set().apply_closure_to_all_completed_buffers(); 5557 5558 DirtyCardQueueSet& dcq = JavaThread::dirty_card_queue_set(); 5559 dcq.merge_bufferlists(&dirty_card_queue_set()); 5560 assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed"); 5561 } 5562 COMPILER2_PRESENT(DerivedPointerTable::update_pointers()); 5563 } 5564 5565 void G1CollectedHeap::free_region_if_empty(HeapRegion* hr, 5566 size_t* pre_used, 5567 FreeRegionList* free_list, 5568 OldRegionSet* old_proxy_set, 5569 HumongousRegionSet* humongous_proxy_set, 5570 HRRSCleanupTask* hrrs_cleanup_task, 5571 bool par) { 5572 if (hr->used() > 0 && hr->max_live_bytes() == 0 && !hr->is_young()) { 5573 if (hr->isHumongous()) { 5574 assert(hr->startsHumongous(), "we should only see starts humongous"); 5575 free_humongous_region(hr, pre_used, free_list, humongous_proxy_set, par); 5576 } else { 5577 _old_set.remove_with_proxy(hr, old_proxy_set); 5578 free_region(hr, pre_used, free_list, par); 5579 } 5580 } else { 5581 hr->rem_set()->do_cleanup_work(hrrs_cleanup_task); 5582 } 5583 } 5584 5585 void G1CollectedHeap::free_region(HeapRegion* hr, 5586 size_t* pre_used, 5587 FreeRegionList* free_list, 5588 bool par) { 5589 assert(!hr->isHumongous(), "this is only for non-humongous regions"); 5590 assert(!hr->is_empty(), "the region should not be empty"); 5591 assert(free_list != NULL, "pre-condition"); 5592 5593 *pre_used += hr->used(); 5594 hr->hr_clear(par, true /* clear_space */); 5595 free_list->add_as_head(hr); 5596 } 5597 5598 void G1CollectedHeap::free_humongous_region(HeapRegion* hr, 5599 size_t* pre_used, 5600 FreeRegionList* free_list, 5601 HumongousRegionSet* humongous_proxy_set, 5602 bool par) { 5603 assert(hr->startsHumongous(), "this is only for starts humongous regions"); 5604 assert(free_list != NULL, "pre-condition"); 5605 assert(humongous_proxy_set != NULL, "pre-condition"); 5606 5607 size_t hr_used = hr->used(); 5608 size_t hr_capacity = hr->capacity(); 5609 size_t hr_pre_used = 0; 5610 _humongous_set.remove_with_proxy(hr, humongous_proxy_set); 5611 hr->set_notHumongous(); 5612 free_region(hr, &hr_pre_used, free_list, par); 5613 5614 uint i = hr->hrs_index() + 1; 5615 uint num = 1; 5616 while (i < n_regions()) { 5617 HeapRegion* curr_hr = region_at(i); 5618 if (!curr_hr->continuesHumongous()) { 5619 break; 5620 } 5621 curr_hr->set_notHumongous(); 5622 free_region(curr_hr, &hr_pre_used, free_list, par); 5623 num += 1; 5624 i += 1; 5625 } 5626 assert(hr_pre_used == hr_used, 5627 err_msg("hr_pre_used: "SIZE_FORMAT" and hr_used: "SIZE_FORMAT" " 5628 "should be the same", hr_pre_used, hr_used)); 5629 *pre_used += hr_pre_used; 5630 } 5631 5632 void G1CollectedHeap::update_sets_after_freeing_regions(size_t pre_used, 5633 FreeRegionList* free_list, 5634 OldRegionSet* old_proxy_set, 5635 HumongousRegionSet* humongous_proxy_set, 5636 bool par) { 5637 if (pre_used > 0) { 5638 Mutex* lock = (par) ? ParGCRareEvent_lock : NULL; 5639 MutexLockerEx x(lock, Mutex::_no_safepoint_check_flag); 5640 assert(_summary_bytes_used >= pre_used, 5641 err_msg("invariant: _summary_bytes_used: "SIZE_FORMAT" " 5642 "should be >= pre_used: "SIZE_FORMAT, 5643 _summary_bytes_used, pre_used)); 5644 _summary_bytes_used -= pre_used; 5645 } 5646 if (free_list != NULL && !free_list->is_empty()) { 5647 MutexLockerEx x(FreeList_lock, Mutex::_no_safepoint_check_flag); 5648 _free_list.add_as_head(free_list); 5649 } 5650 if (old_proxy_set != NULL && !old_proxy_set->is_empty()) { 5651 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5652 _old_set.update_from_proxy(old_proxy_set); 5653 } 5654 if (humongous_proxy_set != NULL && !humongous_proxy_set->is_empty()) { 5655 MutexLockerEx x(OldSets_lock, Mutex::_no_safepoint_check_flag); 5656 _humongous_set.update_from_proxy(humongous_proxy_set); 5657 } 5658 } 5659 5660 class G1ParCleanupCTTask : public AbstractGangTask { 5661 CardTableModRefBS* _ct_bs; 5662 G1CollectedHeap* _g1h; 5663 HeapRegion* volatile _su_head; 5664 public: 5665 G1ParCleanupCTTask(CardTableModRefBS* ct_bs, 5666 G1CollectedHeap* g1h) : 5667 AbstractGangTask("G1 Par Cleanup CT Task"), 5668 _ct_bs(ct_bs), _g1h(g1h) { } 5669 5670 void work(uint worker_id) { 5671 HeapRegion* r; 5672 while (r = _g1h->pop_dirty_cards_region()) { 5673 clear_cards(r); 5674 } 5675 } 5676 5677 void clear_cards(HeapRegion* r) { 5678 // Cards of the survivors should have already been dirtied. 5679 if (!r->is_survivor()) { 5680 _ct_bs->clear(MemRegion(r->bottom(), r->end())); 5681 } 5682 } 5683 }; 5684 5685 #ifndef PRODUCT 5686 class G1VerifyCardTableCleanup: public HeapRegionClosure { 5687 G1CollectedHeap* _g1h; 5688 CardTableModRefBS* _ct_bs; 5689 public: 5690 G1VerifyCardTableCleanup(G1CollectedHeap* g1h, CardTableModRefBS* ct_bs) 5691 : _g1h(g1h), _ct_bs(ct_bs) { } 5692 virtual bool doHeapRegion(HeapRegion* r) { 5693 if (r->is_survivor()) { 5694 _g1h->verify_dirty_region(r); 5695 } else { 5696 _g1h->verify_not_dirty_region(r); 5697 } 5698 return false; 5699 } 5700 }; 5701 5702 void G1CollectedHeap::verify_not_dirty_region(HeapRegion* hr) { 5703 // All of the region should be clean. 5704 CardTableModRefBS* ct_bs = (CardTableModRefBS*)barrier_set(); 5705 MemRegion mr(hr->bottom(), hr->end()); 5706 ct_bs->verify_not_dirty_region(mr); 5707 } 5708 5709 void G1CollectedHeap::verify_dirty_region(HeapRegion* hr) { 5710 // We cannot guarantee that [bottom(),end()] is dirty. Threads 5711 // dirty allocated blocks as they allocate them. The thread that 5712 // retires each region and replaces it with a new one will do a 5713 // maximal allocation to fill in [pre_dummy_top(),end()] but will 5714 // not dirty that area (one less thing to have to do while holding 5715 // a lock). So we can only verify that [bottom(),pre_dummy_top()] 5716 // is dirty. 5717 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5718 MemRegion mr(hr->bottom(), hr->pre_dummy_top()); 5719 ct_bs->verify_dirty_region(mr); 5720 } 5721 5722 void G1CollectedHeap::verify_dirty_young_list(HeapRegion* head) { 5723 CardTableModRefBS* ct_bs = (CardTableModRefBS*) barrier_set(); 5724 for (HeapRegion* hr = head; hr != NULL; hr = hr->get_next_young_region()) { 5725 verify_dirty_region(hr); 5726 } 5727 } 5728 5729 void G1CollectedHeap::verify_dirty_young_regions() { 5730 verify_dirty_young_list(_young_list->first_region()); 5731 verify_dirty_young_list(_young_list->first_survivor_region()); 5732 } 5733 #endif 5734 5735 void G1CollectedHeap::cleanUpCardTable() { 5736 CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set()); 5737 double start = os::elapsedTime(); 5738 5739 { 5740 // Iterate over the dirty cards region list. 5741 G1ParCleanupCTTask cleanup_task(ct_bs, this); 5742 5743 if (G1CollectedHeap::use_parallel_gc_threads()) { 5744 set_par_threads(); 5745 workers()->run_task(&cleanup_task); 5746 set_par_threads(0); 5747 } else { 5748 while (_dirty_cards_region_list) { 5749 HeapRegion* r = _dirty_cards_region_list; 5750 cleanup_task.clear_cards(r); 5751 _dirty_cards_region_list = r->get_next_dirty_cards_region(); 5752 if (_dirty_cards_region_list == r) { 5753 // The last region. 5754 _dirty_cards_region_list = NULL; 5755 } 5756 r->set_next_dirty_cards_region(NULL); 5757 } 5758 } 5759 #ifndef PRODUCT 5760 if (G1VerifyCTCleanup || VerifyAfterGC) { 5761 G1VerifyCardTableCleanup cleanup_verifier(this, ct_bs); 5762 heap_region_iterate(&cleanup_verifier); 5763 } 5764 #endif 5765 } 5766 5767 double elapsed = os::elapsedTime() - start; 5768 g1_policy()->phase_times()->record_clear_ct_time(elapsed * 1000.0); 5769 } 5770 5771 void G1CollectedHeap::free_collection_set(HeapRegion* cs_head) { 5772 size_t pre_used = 0; 5773 FreeRegionList local_free_list("Local List for CSet Freeing"); 5774 5775 double young_time_ms = 0.0; 5776 double non_young_time_ms = 0.0; 5777 5778 // Since the collection set is a superset of the the young list, 5779 // all we need to do to clear the young list is clear its 5780 // head and length, and unlink any young regions in the code below 5781 _young_list->clear(); 5782 5783 G1CollectorPolicy* policy = g1_policy(); 5784 5785 double start_sec = os::elapsedTime(); 5786 bool non_young = true; 5787 5788 HeapRegion* cur = cs_head; 5789 int age_bound = -1; 5790 size_t rs_lengths = 0; 5791 5792 while (cur != NULL) { 5793 assert(!is_on_master_free_list(cur), "sanity"); 5794 if (non_young) { 5795 if (cur->is_young()) { 5796 double end_sec = os::elapsedTime(); 5797 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5798 non_young_time_ms += elapsed_ms; 5799 5800 start_sec = os::elapsedTime(); 5801 non_young = false; 5802 } 5803 } else { 5804 if (!cur->is_young()) { 5805 double end_sec = os::elapsedTime(); 5806 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5807 young_time_ms += elapsed_ms; 5808 5809 start_sec = os::elapsedTime(); 5810 non_young = true; 5811 } 5812 } 5813 5814 rs_lengths += cur->rem_set()->occupied(); 5815 5816 HeapRegion* next = cur->next_in_collection_set(); 5817 assert(cur->in_collection_set(), "bad CS"); 5818 cur->set_next_in_collection_set(NULL); 5819 cur->set_in_collection_set(false); 5820 5821 if (cur->is_young()) { 5822 int index = cur->young_index_in_cset(); 5823 assert(index != -1, "invariant"); 5824 assert((uint) index < policy->young_cset_region_length(), "invariant"); 5825 size_t words_survived = _surviving_young_words[index]; 5826 cur->record_surv_words_in_group(words_survived); 5827 5828 // At this point the we have 'popped' cur from the collection set 5829 // (linked via next_in_collection_set()) but it is still in the 5830 // young list (linked via next_young_region()). Clear the 5831 // _next_young_region field. 5832 cur->set_next_young_region(NULL); 5833 } else { 5834 int index = cur->young_index_in_cset(); 5835 assert(index == -1, "invariant"); 5836 } 5837 5838 assert( (cur->is_young() && cur->young_index_in_cset() > -1) || 5839 (!cur->is_young() && cur->young_index_in_cset() == -1), 5840 "invariant" ); 5841 5842 if (!cur->evacuation_failed()) { 5843 MemRegion used_mr = cur->used_region(); 5844 5845 // And the region is empty. 5846 assert(!used_mr.is_empty(), "Should not have empty regions in a CS."); 5847 free_region(cur, &pre_used, &local_free_list, false /* par */); 5848 } else { 5849 cur->uninstall_surv_rate_group(); 5850 if (cur->is_young()) { 5851 cur->set_young_index_in_cset(-1); 5852 } 5853 cur->set_not_young(); 5854 cur->set_evacuation_failed(false); 5855 // The region is now considered to be old. 5856 _old_set.add(cur); 5857 } 5858 cur = next; 5859 } 5860 5861 policy->record_max_rs_lengths(rs_lengths); 5862 policy->cset_regions_freed(); 5863 5864 double end_sec = os::elapsedTime(); 5865 double elapsed_ms = (end_sec - start_sec) * 1000.0; 5866 5867 if (non_young) { 5868 non_young_time_ms += elapsed_ms; 5869 } else { 5870 young_time_ms += elapsed_ms; 5871 } 5872 5873 update_sets_after_freeing_regions(pre_used, &local_free_list, 5874 NULL /* old_proxy_set */, 5875 NULL /* humongous_proxy_set */, 5876 false /* par */); 5877 policy->phase_times()->record_young_free_cset_time_ms(young_time_ms); 5878 policy->phase_times()->record_non_young_free_cset_time_ms(non_young_time_ms); 5879 } 5880 5881 // This routine is similar to the above but does not record 5882 // any policy statistics or update free lists; we are abandoning 5883 // the current incremental collection set in preparation of a 5884 // full collection. After the full GC we will start to build up 5885 // the incremental collection set again. 5886 // This is only called when we're doing a full collection 5887 // and is immediately followed by the tearing down of the young list. 5888 5889 void G1CollectedHeap::abandon_collection_set(HeapRegion* cs_head) { 5890 HeapRegion* cur = cs_head; 5891 5892 while (cur != NULL) { 5893 HeapRegion* next = cur->next_in_collection_set(); 5894 assert(cur->in_collection_set(), "bad CS"); 5895 cur->set_next_in_collection_set(NULL); 5896 cur->set_in_collection_set(false); 5897 cur->set_young_index_in_cset(-1); 5898 cur = next; 5899 } 5900 } 5901 5902 void G1CollectedHeap::set_free_regions_coming() { 5903 if (G1ConcRegionFreeingVerbose) { 5904 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5905 "setting free regions coming"); 5906 } 5907 5908 assert(!free_regions_coming(), "pre-condition"); 5909 _free_regions_coming = true; 5910 } 5911 5912 void G1CollectedHeap::reset_free_regions_coming() { 5913 assert(free_regions_coming(), "pre-condition"); 5914 5915 { 5916 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5917 _free_regions_coming = false; 5918 SecondaryFreeList_lock->notify_all(); 5919 } 5920 5921 if (G1ConcRegionFreeingVerbose) { 5922 gclog_or_tty->print_cr("G1ConcRegionFreeing [cm thread] : " 5923 "reset free regions coming"); 5924 } 5925 } 5926 5927 void G1CollectedHeap::wait_while_free_regions_coming() { 5928 // Most of the time we won't have to wait, so let's do a quick test 5929 // first before we take the lock. 5930 if (!free_regions_coming()) { 5931 return; 5932 } 5933 5934 if (G1ConcRegionFreeingVerbose) { 5935 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5936 "waiting for free regions"); 5937 } 5938 5939 { 5940 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 5941 while (free_regions_coming()) { 5942 SecondaryFreeList_lock->wait(Mutex::_no_safepoint_check_flag); 5943 } 5944 } 5945 5946 if (G1ConcRegionFreeingVerbose) { 5947 gclog_or_tty->print_cr("G1ConcRegionFreeing [other] : " 5948 "done waiting for free regions"); 5949 } 5950 } 5951 5952 void G1CollectedHeap::set_region_short_lived_locked(HeapRegion* hr) { 5953 assert(heap_lock_held_for_gc(), 5954 "the heap lock should already be held by or for this thread"); 5955 _young_list->push_region(hr); 5956 } 5957 5958 class NoYoungRegionsClosure: public HeapRegionClosure { 5959 private: 5960 bool _success; 5961 public: 5962 NoYoungRegionsClosure() : _success(true) { } 5963 bool doHeapRegion(HeapRegion* r) { 5964 if (r->is_young()) { 5965 gclog_or_tty->print_cr("Region ["PTR_FORMAT", "PTR_FORMAT") tagged as young", 5966 r->bottom(), r->end()); 5967 _success = false; 5968 } 5969 return false; 5970 } 5971 bool success() { return _success; } 5972 }; 5973 5974 bool G1CollectedHeap::check_young_list_empty(bool check_heap, bool check_sample) { 5975 bool ret = _young_list->check_list_empty(check_sample); 5976 5977 if (check_heap) { 5978 NoYoungRegionsClosure closure; 5979 heap_region_iterate(&closure); 5980 ret = ret && closure.success(); 5981 } 5982 5983 return ret; 5984 } 5985 5986 class TearDownRegionSetsClosure : public HeapRegionClosure { 5987 private: 5988 OldRegionSet *_old_set; 5989 5990 public: 5991 TearDownRegionSetsClosure(OldRegionSet* old_set) : _old_set(old_set) { } 5992 5993 bool doHeapRegion(HeapRegion* r) { 5994 if (r->is_empty()) { 5995 // We ignore empty regions, we'll empty the free list afterwards 5996 } else if (r->is_young()) { 5997 // We ignore young regions, we'll empty the young list afterwards 5998 } else if (r->isHumongous()) { 5999 // We ignore humongous regions, we're not tearing down the 6000 // humongous region set 6001 } else { 6002 // The rest should be old 6003 _old_set->remove(r); 6004 } 6005 return false; 6006 } 6007 6008 ~TearDownRegionSetsClosure() { 6009 assert(_old_set->is_empty(), "post-condition"); 6010 } 6011 }; 6012 6013 void G1CollectedHeap::tear_down_region_sets(bool free_list_only) { 6014 assert_at_safepoint(true /* should_be_vm_thread */); 6015 6016 if (!free_list_only) { 6017 TearDownRegionSetsClosure cl(&_old_set); 6018 heap_region_iterate(&cl); 6019 6020 // Need to do this after the heap iteration to be able to 6021 // recognize the young regions and ignore them during the iteration. 6022 _young_list->empty_list(); 6023 } 6024 _free_list.remove_all(); 6025 } 6026 6027 class RebuildRegionSetsClosure : public HeapRegionClosure { 6028 private: 6029 bool _free_list_only; 6030 OldRegionSet* _old_set; 6031 FreeRegionList* _free_list; 6032 size_t _total_used; 6033 6034 public: 6035 RebuildRegionSetsClosure(bool free_list_only, 6036 OldRegionSet* old_set, FreeRegionList* free_list) : 6037 _free_list_only(free_list_only), 6038 _old_set(old_set), _free_list(free_list), _total_used(0) { 6039 assert(_free_list->is_empty(), "pre-condition"); 6040 if (!free_list_only) { 6041 assert(_old_set->is_empty(), "pre-condition"); 6042 } 6043 } 6044 6045 bool doHeapRegion(HeapRegion* r) { 6046 if (r->continuesHumongous()) { 6047 return false; 6048 } 6049 6050 if (r->is_empty()) { 6051 // Add free regions to the free list 6052 _free_list->add_as_tail(r); 6053 } else if (!_free_list_only) { 6054 assert(!r->is_young(), "we should not come across young regions"); 6055 6056 if (r->isHumongous()) { 6057 // We ignore humongous regions, we left the humongous set unchanged 6058 } else { 6059 // The rest should be old, add them to the old set 6060 _old_set->add(r); 6061 } 6062 _total_used += r->used(); 6063 } 6064 6065 return false; 6066 } 6067 6068 size_t total_used() { 6069 return _total_used; 6070 } 6071 }; 6072 6073 void G1CollectedHeap::rebuild_region_sets(bool free_list_only) { 6074 assert_at_safepoint(true /* should_be_vm_thread */); 6075 6076 RebuildRegionSetsClosure cl(free_list_only, &_old_set, &_free_list); 6077 heap_region_iterate(&cl); 6078 6079 if (!free_list_only) { 6080 _summary_bytes_used = cl.total_used(); 6081 } 6082 assert(_summary_bytes_used == recalculate_used(), 6083 err_msg("inconsistent _summary_bytes_used, " 6084 "value: "SIZE_FORMAT" recalculated: "SIZE_FORMAT, 6085 _summary_bytes_used, recalculate_used())); 6086 } 6087 6088 void G1CollectedHeap::set_refine_cte_cl_concurrency(bool concurrent) { 6089 _refine_cte_cl->set_concurrent(concurrent); 6090 } 6091 6092 bool G1CollectedHeap::is_in_closed_subset(const void* p) const { 6093 HeapRegion* hr = heap_region_containing(p); 6094 if (hr == NULL) { 6095 return is_in_permanent(p); 6096 } else { 6097 return hr->is_in(p); 6098 } 6099 } 6100 6101 // Methods for the mutator alloc region 6102 6103 HeapRegion* G1CollectedHeap::new_mutator_alloc_region(size_t word_size, 6104 bool force) { 6105 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6106 assert(!force || g1_policy()->can_expand_young_list(), 6107 "if force is true we should be able to expand the young list"); 6108 bool young_list_full = g1_policy()->is_young_list_full(); 6109 if (force || !young_list_full) { 6110 HeapRegion* new_alloc_region = new_region(word_size, 6111 false /* do_expand */); 6112 if (new_alloc_region != NULL) { 6113 set_region_short_lived_locked(new_alloc_region); 6114 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Eden, young_list_full); 6115 return new_alloc_region; 6116 } 6117 } 6118 return NULL; 6119 } 6120 6121 void G1CollectedHeap::retire_mutator_alloc_region(HeapRegion* alloc_region, 6122 size_t allocated_bytes) { 6123 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6124 assert(alloc_region->is_young(), "all mutator alloc regions should be young"); 6125 6126 g1_policy()->add_region_to_incremental_cset_lhs(alloc_region); 6127 _summary_bytes_used += allocated_bytes; 6128 _hr_printer.retire(alloc_region); 6129 // We update the eden sizes here, when the region is retired, 6130 // instead of when it's allocated, since this is the point that its 6131 // used space has been recored in _summary_bytes_used. 6132 g1mm()->update_eden_size(); 6133 } 6134 6135 HeapRegion* MutatorAllocRegion::allocate_new_region(size_t word_size, 6136 bool force) { 6137 return _g1h->new_mutator_alloc_region(word_size, force); 6138 } 6139 6140 void G1CollectedHeap::set_par_threads() { 6141 // Don't change the number of workers. Use the value previously set 6142 // in the workgroup. 6143 assert(G1CollectedHeap::use_parallel_gc_threads(), "shouldn't be here otherwise"); 6144 uint n_workers = workers()->active_workers(); 6145 assert(UseDynamicNumberOfGCThreads || 6146 n_workers == workers()->total_workers(), 6147 "Otherwise should be using the total number of workers"); 6148 if (n_workers == 0) { 6149 assert(false, "Should have been set in prior evacuation pause."); 6150 n_workers = ParallelGCThreads; 6151 workers()->set_active_workers(n_workers); 6152 } 6153 set_par_threads(n_workers); 6154 } 6155 6156 void MutatorAllocRegion::retire_region(HeapRegion* alloc_region, 6157 size_t allocated_bytes) { 6158 _g1h->retire_mutator_alloc_region(alloc_region, allocated_bytes); 6159 } 6160 6161 // Methods for the GC alloc regions 6162 6163 HeapRegion* G1CollectedHeap::new_gc_alloc_region(size_t word_size, 6164 uint count, 6165 GCAllocPurpose ap) { 6166 assert(FreeList_lock->owned_by_self(), "pre-condition"); 6167 6168 if (count < g1_policy()->max_regions(ap)) { 6169 HeapRegion* new_alloc_region = new_region(word_size, 6170 true /* do_expand */); 6171 if (new_alloc_region != NULL) { 6172 // We really only need to do this for old regions given that we 6173 // should never scan survivors. But it doesn't hurt to do it 6174 // for survivors too. 6175 new_alloc_region->set_saved_mark(); 6176 if (ap == GCAllocForSurvived) { 6177 new_alloc_region->set_survivor(); 6178 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Survivor); 6179 } else { 6180 _hr_printer.alloc(new_alloc_region, G1HRPrinter::Old); 6181 } 6182 bool during_im = g1_policy()->during_initial_mark_pause(); 6183 new_alloc_region->note_start_of_copying(during_im); 6184 return new_alloc_region; 6185 } else { 6186 g1_policy()->note_alloc_region_limit_reached(ap); 6187 } 6188 } 6189 return NULL; 6190 } 6191 6192 void G1CollectedHeap::retire_gc_alloc_region(HeapRegion* alloc_region, 6193 size_t allocated_bytes, 6194 GCAllocPurpose ap) { 6195 bool during_im = g1_policy()->during_initial_mark_pause(); 6196 alloc_region->note_end_of_copying(during_im); 6197 g1_policy()->record_bytes_copied_during_gc(allocated_bytes); 6198 if (ap == GCAllocForSurvived) { 6199 young_list()->add_survivor_region(alloc_region); 6200 } else { 6201 _old_set.add(alloc_region); 6202 } 6203 _hr_printer.retire(alloc_region); 6204 } 6205 6206 HeapRegion* SurvivorGCAllocRegion::allocate_new_region(size_t word_size, 6207 bool force) { 6208 assert(!force, "not supported for GC alloc regions"); 6209 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForSurvived); 6210 } 6211 6212 void SurvivorGCAllocRegion::retire_region(HeapRegion* alloc_region, 6213 size_t allocated_bytes) { 6214 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6215 GCAllocForSurvived); 6216 } 6217 6218 HeapRegion* OldGCAllocRegion::allocate_new_region(size_t word_size, 6219 bool force) { 6220 assert(!force, "not supported for GC alloc regions"); 6221 return _g1h->new_gc_alloc_region(word_size, count(), GCAllocForTenured); 6222 } 6223 6224 void OldGCAllocRegion::retire_region(HeapRegion* alloc_region, 6225 size_t allocated_bytes) { 6226 _g1h->retire_gc_alloc_region(alloc_region, allocated_bytes, 6227 GCAllocForTenured); 6228 } 6229 // Heap region set verification 6230 6231 class VerifyRegionListsClosure : public HeapRegionClosure { 6232 private: 6233 FreeRegionList* _free_list; 6234 OldRegionSet* _old_set; 6235 HumongousRegionSet* _humongous_set; 6236 uint _region_count; 6237 6238 public: 6239 VerifyRegionListsClosure(OldRegionSet* old_set, 6240 HumongousRegionSet* humongous_set, 6241 FreeRegionList* free_list) : 6242 _old_set(old_set), _humongous_set(humongous_set), 6243 _free_list(free_list), _region_count(0) { } 6244 6245 uint region_count() { return _region_count; } 6246 6247 bool doHeapRegion(HeapRegion* hr) { 6248 _region_count += 1; 6249 6250 if (hr->continuesHumongous()) { 6251 return false; 6252 } 6253 6254 if (hr->is_young()) { 6255 // TODO 6256 } else if (hr->startsHumongous()) { 6257 _humongous_set->verify_next_region(hr); 6258 } else if (hr->is_empty()) { 6259 _free_list->verify_next_region(hr); 6260 } else { 6261 _old_set->verify_next_region(hr); 6262 } 6263 return false; 6264 } 6265 }; 6266 6267 HeapRegion* G1CollectedHeap::new_heap_region(uint hrs_index, 6268 HeapWord* bottom) { 6269 HeapWord* end = bottom + HeapRegion::GrainWords; 6270 MemRegion mr(bottom, end); 6271 assert(_g1_reserved.contains(mr), "invariant"); 6272 // This might return NULL if the allocation fails 6273 return new HeapRegion(hrs_index, _bot_shared, mr, true /* is_zeroed */); 6274 } 6275 6276 void G1CollectedHeap::verify_region_sets() { 6277 assert_heap_locked_or_at_safepoint(true /* should_be_vm_thread */); 6278 6279 // First, check the explicit lists. 6280 _free_list.verify(); 6281 { 6282 // Given that a concurrent operation might be adding regions to 6283 // the secondary free list we have to take the lock before 6284 // verifying it. 6285 MutexLockerEx x(SecondaryFreeList_lock, Mutex::_no_safepoint_check_flag); 6286 _secondary_free_list.verify(); 6287 } 6288 _old_set.verify(); 6289 _humongous_set.verify(); 6290 6291 // If a concurrent region freeing operation is in progress it will 6292 // be difficult to correctly attributed any free regions we come 6293 // across to the correct free list given that they might belong to 6294 // one of several (free_list, secondary_free_list, any local lists, 6295 // etc.). So, if that's the case we will skip the rest of the 6296 // verification operation. Alternatively, waiting for the concurrent 6297 // operation to complete will have a non-trivial effect on the GC's 6298 // operation (no concurrent operation will last longer than the 6299 // interval between two calls to verification) and it might hide 6300 // any issues that we would like to catch during testing. 6301 if (free_regions_coming()) { 6302 return; 6303 } 6304 6305 // Make sure we append the secondary_free_list on the free_list so 6306 // that all free regions we will come across can be safely 6307 // attributed to the free_list. 6308 append_secondary_free_list_if_not_empty_with_lock(); 6309 6310 // Finally, make sure that the region accounting in the lists is 6311 // consistent with what we see in the heap. 6312 _old_set.verify_start(); 6313 _humongous_set.verify_start(); 6314 _free_list.verify_start(); 6315 6316 VerifyRegionListsClosure cl(&_old_set, &_humongous_set, &_free_list); 6317 heap_region_iterate(&cl); 6318 6319 _old_set.verify_end(); 6320 _humongous_set.verify_end(); 6321 _free_list.verify_end(); 6322 }